Constructive approaches to the rigidity of frameworks Viet Hang Nguyen
To cite this version: Viet Hang Nguyen. Constructive approaches to the rigidity of frameworks. Math´ematiques g´en´erales[math.GM]. Universit´ede Grenoble, 2013. Fran¸cais.
HAL Id: tel-00992387 https://tel.archives-ouvertes.fr/tel-00992387 Submitted on 17 May 2014
HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est archive for the deposit and dissemination of sci- destin´eeau d´epˆotet `ala diffusion de documents entific research documents, whether they are pub- scientifiques de niveau recherche, publi´esou non, lished or not. The documents may come from ´emanant des ´etablissements d’enseignement et de teaching and research institutions in France or recherche fran¸caisou ´etrangers,des laboratoires abroad, or from public or private research centers. publics ou priv´es. THÈSE Pour obtenir le grade de DOCTEUR DE L’UNIVERSITÉ DE GRENOBLE Spécialité : Mathématique et Informatique
Arrêté ministériel : 7 aoutˆ 2006
Présentée par Việt Hằng Nguyễn
Thèse dirigée par Zoltán Szigeti préparée au sein du Laboratoire G-SCOP et de l’École Doctorale MSTII
Constructive approaches to the rigidity of frameworks
Thèse soutenue publiquement le 17 octobre 2013, devant le jury composé de :
M. Jorge Ramírez Alfonsín Professeur, Université Montpellier 2, Président M. András Recski Professeur, Budapest University of Technology and Economics, Rapporteur M. Walter Whiteley Professeur, York University (Toronto), Rapporteur M. Victor Chepoi Professeur, Université de la Mediterranée (Marseille), Examinateur M. Tibor Jordán Professeur, Eotv¨ os¨ Loránd University (Budapest), Examinateur M. Zoltán Szigeti Professeur, Grenoble INP, Laboratoire G-SCOP, Directeur de thèse Acknowledgements
I would like to thank the members of the committee for their comments, suggestions on the manuscript as well as inspiring questions. I would like to express my deepest thank to my advisor Zolt´an Szigeti for his constant support, encouragement, counsel and especially the great freedom I enjoyed during these three years. My special thank goes to Andr´as Seb¨ofor his support and also for his energy which propagates in our research group and motivates us all. I am deeply grateful to Olivier Durand de Gevigney, my academic brother, co-author and true friend. I greatly appreciate his help and all the time we have shared. I would like also to thank my other collaborators Abdo Alfakih, Bill Jackson and Tibor Jord´an for the joint works and many things I learned from them. I am thankful to all the people in our research group for discussions and sem- inars. I thank friends and staffs in the G-SCOP laboratory for having created a pleasant atmosphere. I would like to express my thank to my family in Vietnam for their love and encouragement. I am deeply grateful to my family-in-law and especially to my mother-in-law for her kindness and help. There are no words to express my thank to my husband Guillaume who has been and will be always with me in our journey of life. I am also thankful to our little angel Daniel Hai Phong, our unending source of joy and inspiration.
1 2 Contents
1 Introduction 7 1.1 Frameworks and rigidity ...... 10 1.2 Our contributions and organization of the thesis ...... 15
2 Preliminaries 19 2.1 Basic notions ...... 20 2.2 Graphs and digraphs ...... 20 2.3 Matroid theory ...... 23 2.4 Submodular functions ...... 26 2.5 Algebra and linear algebra ...... 27 2.6 Congruent motions and motions of rigid bodies ...... 30
3 Rigidity theory 33 3.1 Introduction ...... 34 3.2 Various types of rigidity ...... 35 3.2.1 Local rigidity ...... 35 3.2.2 Infinitesimal rigidity and rigidity matrix ...... 36 3.2.3 Local rigidity versus infinitesimal rigidity ...... 38 3.2.4 Static rigidity ...... 40 3.2.5 Infinitesimal rigidity versus static rigidity ...... 41 3.2.6 Global rigidity ...... 41 3.2.7 Universal rigidity ...... 42 3.2.8 Stress matrices ...... 42 3.3 Combinatorial rigidity ...... 44
3 Contents
4 Matroid approach 51 4.1 Introduction ...... 52 4.2 Characterizing abstract rigidity matroids ...... 53 4.3 1-extendable abstract rigidity matroids ...... 54 4.4 Submodular functions inducing ARMs ...... 58 4.5 A potential application ...... 61
5 Inductive constructions and decompositions 63 5.1 Introduction ...... 65 5.2 Graded sparse graphs ...... 68 5.2.1 Introduction ...... 68 5.2.2 A reduction theorem for graded sparse graphs ...... 70 5.2.3 Inductive construction of graded sparse graphs ...... 75 5.2.4 Decomposition of graded sparse graphs ...... 75 5.2.5 Graded sparse matroids ...... 81 5.3 (b, l)-sparse graphs ...... 86 5.3.1 Inductive construction ...... 86 5.3.2 (b, l)-pebble games ...... 93 5.4 Packing of matroid-based arborescences ...... 96 5.4.1 Introduction ...... 96 5.4.2 Proof of the main theorem ...... 102 5.4.3 Polyhedral description ...... 105 5.4.4 Algorithmic aspects ...... 106 5.4.5 Further remarks ...... 107
6 Infinitesimal rigidity 109 6.1 Introduction ...... 110 6.2 Body-bar frameworks with boundary ...... 112 6.2.1 Body-bar frameworks ...... 112 6.2.2 Body-bar frameworks with bar-boundary ...... 115 6.3 Body-length-direction frameworks ...... 119 6.3.1 Introduction ...... 119 6.3.2 Frameworks with length-direction-rigid bodies ...... 120 6.3.3 Frameworks with direction-rigid bodies ...... 122 6.3.4 Frameworks with length-rigid bodies ...... 126 6.4 Direction-length frameworks ...... 131 6.4.1 Introduction ...... 131
4 Contents
6.4.2 0-extensions for direction-length graphs ...... 133 6.4.3 1-extensions for direction-length graphs ...... 135 6.4.4 Union of two spanning trees ...... 141
7 Global rigidity of direction-length frameworks 143 7.1 Introduction ...... 144 7.2 Quasi-generic direction-length frameworks ...... 145 7.3 Proof of Theorem 7.1.2 ...... 147 7.4 Further remarks ...... 156
8 Universal rigidity 159 8.1 Introduction ...... 160 8.2 Universal rigidity in R1 ...... 161 8.2.1 Universal rigidity of bipartite frameworks ...... 161 8.2.2 Bipartite graphs in general dimension ...... 164 8.2.3 Observations, open questions, conjectures ...... 165 8.3 Universal rigidity of tensegrity frameworks ...... 172 8.3.1 Introduction ...... 172 8.3.2 Preliminaries ...... 175 8.3.3 The configuration matrix and Gale matrices ...... 177 8.3.4 Dominated tensegrity frameworks ...... 178 8.3.5 Affinely dominated tensegrity frameworks ...... 180 8.3.6 Proof of main results ...... 183 8.3.7 Counter-example and conjectures ...... 185
Conclusion 189 References ...... 193
5 Contents
6 Chapter 1
Introduction
Contents 1.1 Frameworks and rigidity ...... 10 1.2 Our contributions and organization of the thesis . . . 15
7 La th´eorie de la rigidit´ese pr´eoccupe originellement de la rigidit´e/flexibilit´edes charpentes. Une charpente est un mod`ele math´ematique d´ecrivant une structure r´eelle. Une des premi`eres sources de structures qui ont suscit´el’´etude de la rigidit´e vient de l’ing´enierie des structures. Consid´erons une structure planaire constitu´ee de barres solides jointes aux extr´emit´es par des jonctions permettant `adeux barres jointes de bouger librement dans le plan autour du joint. Une question naturelle est de savoir si la structure peut ˆetre d´eform´ee. Par exemple, la structure planaire rectangulaire de la Figure 1.1 peut ˆetre d´eform´ee continˆument en une structure planaire non-rectangulaire comme repr´esent´epar la figure en pointill´es. Nous dis- ons que la structure est flexible. Au contraire, la structure planaire de la Figure 1.2 ne peut pas ˆetre d´eform´ee continˆument en une autre forme sans ˆetre cass´ee. Nous disons que la structure est rigide, ou plus pr´ecis´ement, localement rigide. Ce type de structures barres-et-joints planaires peut ˆetre mod´elis´ee par une paire (G, p) form´ee d’un graphe G = (V,E) et d’une application p : V → R2. Les sommets V de G correspondent aux joints et les arˆetes E de G correspondent aux barres. L’application p d´etermine la position des joints. De la mˆeme mani`ere, une structure barres-et-joints en dimension 3 (et plus g´en´eralement en dimension d) peut ˆetre mod´elis´ee par une application p : V → R3 (et p : V → Rd, respective- ment). Nous appelons cette paire une charpente barres-et-joints. Un d´eplacement continu d’une charpente (G, p) dans Rd est un d´eplacement continu des sommets de G dans Rd qui conserve les longueurs des arˆetes. Une charpente est dite locale- ment rigide si tous les d´eplacements continus de (G, p) pr´eservent aussi la distance entre chaque paire de sommets de G. De mani`ere ´equivalente, tout d´eplacement continu de (G, p) est la restriction d’un d´eplacement congruent de l’espace tout entier. D´eterminer si une charpente est localement rigide est difficile en g´en´eral. Par contre, on peut s’int´eresser aux d´eplacements instantan´es des joints du probl`eme lin´earis´e. Un d´eplacement infinit´esimal d’un vecteur µ(v) `achaque sommet v ∈ V (µ(v) peut ˆetre vu comme la vitesse instantan´ee du joint) tel que pour chaque arˆete (i.e. barre) les d´eplacements instantan´es le long de cette arˆete induits par ce d´eplacement infinit´esimal aux deux extr´emit´es sont identiques, ce qui signifie
(p(u) − p(v))(µ(u) − µ(v))=0 pourtout uv ∈ E. (1.0.1)
(Voir Figure 3.2.) Une charpente est dite infinit´esimalement rigide si tous ses d´eplacements infinit´esimaux peuvent ˆetre obtenus `apartir des vecteurs de vitesse instantan´ee de la restriction d’un d´eplacement congruent de tout l’espace sur V .
8 Chapter 1. Introduction
En fait, la rigidit´einfinit´esimale est une propri´et´eplus forte que la rigidit´elocale, mais c’est une bonne alternative `ala rigidit´elocale car il s’av`ere qu’elles co¨ıncident dans la plupart des cas, `asavoir pour les charpentes en position g´en´erique. La rigidit´einfinit´esimale offre plus de prise que la rigidit´elocale puisque elle peut ˆetre d´etermin´ee par le calcul du rang de la matrice d´eriv´ee du syst`eme lin´eaire (1.1.1) pour µ. L’´etude de la rigidit´ea trouv´edes applications pour pr´edire la flexibilit´ede prot´eines. D’autres notions de rigidit´em´eritent notre attention. La premi`ere se pr´eoccupe de l’unicit´eglobale des charpentes, `asavoir, si l’ensemble des contraintes de dis- tance impos´ees aux barres suffit `ad´eterminer la charpente modulo une congruence. L’´etude de la rigidit´eglobale a des applications dans la localisation dans des r´eseaux de capteurs. Bien que la rigidit´eglobale apparaˆıttr`es diff´erente de la rigidit´elocale, elles sont en fait ´etroitement li´ees. Il est facile de voir que la rigidit´eglobale implique la rigidit´elocale. Dans l’autre direction, il est d´emontr´equ’une charpente planaire g´en´erique est globalement rigide si et seulement si il faut enlever au moins 3 som- mets pour la d´econnecter et elle reste localement rigide apr`es la suppression d’une arˆete quelconque. En pratique, pour calculer la position des capteurs, on peut utiliser des algo- rithmes bas´es sur la programmation semi-d´efinie. Cependant, mˆeme lorsque la position des capteurs est unique dans l’espace (plan ou en dimension 3), ces algorithmes peuvent trouver une configuration en dimension sup´erieure. Pour ´eviter cet effet ind´esirable, la rigidit´e universelle, une notion plus forte que la rigidit´eglobale est requise. Une charpente est universelle- ment rigide si elle est exclusivement d´etermin´ee modulo une congruence dans tous les espaces. Caract´eriser combinatoirement la rigidit´euniverselle des charpentes g´en´eriques semble plus difficile que la rigidit´eglobale, en partie car elle ne d´epend pas uniquement du graphe sous-jacent, mˆeme lorsque les sommets sont dans une position g´en´erique. En fait, mˆeme en dimension 1, aucune caract´erisation combi- natoire de la rigidit´euniverselle est connue.
9 1.1. Frameworks and rigidity
1.1 Frameworks and rigidity
Rigidity theory originally concerns with the rigidity/flexibility of frameworks. A framework is a mathematical model describing a real structure. An early source of structures that provoked the study of rigidity is from architectural engineering. Consider a planar structure constituting of solid bars that are joined at extremities, called joints, by junctions that allow two joined bars to move freely in the plane about this joint. A natural question is whether the shape of the structure can be deformed. For example, the rectangular planar structure in Figure 1.1 can be deformed continuously to a non-rectangular one as shown by the dashed figure. We say that it is flexible. Meanwhile, the planar structure in Figure 1.2 can not be deformed continuously to another shape without being ripped apart. We say that it is rigid, or more precisely, locally rigid. •• ••
• • • • Figure 1.1: A flexible bar-and-joint structure in the plane.
••
• •
Figure 1.2: A rigid bar-and-joint structure in the plane.
Such a planar bar-and-joint structure can be modeled by a pair (G, p) of a graph G = (V,E) and a map p : V → R2. The vertices V of G correspond to the joints and the edges E of G correspond to the bars. The map p determines the location of the joints. In the same way, a bar-and-joint structure in 3-space (and more generally in dimension d) can be modeled with a map p : V → R3 (and p : V → Rd, respectively). We call this pair a bar-joint framework. A continuous motion of a framework (G, p) in Rd is a continuous motion of the vertices of G in Rd which preserves the edge lengths. A framework is said to be locally rigid if every continuous motion of (G, p) preserves also the pairwise distance between all
10 Chapter 1. Introduction
• • • • • • • • • • • • • • • • • • a) b)
Figure 1.3: a) A molecule structure with atoms and bonds. b) The corresponding bar-joint framework. vertices of G. Equivalently, every continuous motion of (G, p) is the restriction of a congruent motion of the whole space. The problem of determining the local rigidity of a framework is hard in general. One way to deal with this is to consider the linearized problem where we focus on the instantaneous motion of the joints. An infinitesimal motion of a framework is an assignment of a vector µ(v) to each vertex v ∈ V (µ(v) can be viewed as the instantaneous velocity of the joint) such that for each edge (i.e., bar) the instantaneous displacements along this edge induced by this infinitesimal motion at two ends are the same, which means
(p(u) − p(v))(µ(u) − µ(v)) = 0 for all uv ∈ E. (1.1.1)
(See Figure 3.2.) A framework is said to be infinitesimally rigid if all its infinitesimal motions can be obtained as the instantaneous velocity vectors of a restriction of a congruent motion of the whole space on V . As a matter of fact, infinitesimal rigidity is a stronger property than local rigidity, but it is a good alternative for local rigidity since it turns out that they coincide in “most” cases, namely for frameworks in generic position. Infinitesimal rigidity is more tractable than local rigidity since it can be decided by calculating the rank of a matrix derived from the linear system (1.1.1) for µ. The study of local rigidity/flexibility has found important applications in pre- dicting flexibility of protein molecules. We can regard a molecular structure with covalent bonds as a bar-joint framework in 3-space: each atom is a joint and each covalent bond works as a bar that fixes the distance between two atoms. Since the angles between convalent bonds of an atom are also fixed, we need to add one more edge between each pair of neighbors of an atom (Figure 1.3). As protein molecules
11 1.1. Frameworks and rigidity are often very large with thousands of atoms and bonds, the linear algebra approach by calculating the rank of the rigidity matrix becomes non efficient. The study of large frameworks necessitates the combinatorial results on the underlying graphs of these frameworks which would facilitate the design of efficient algorithms. The fundamental result by Laman (1970) asserts that the local/infinitesimal rigidity of a generic framework in the plane can be discerned by a simple counting condition on vertices and edges. This counting condition is based on the following intuitive reasoning: Every motion of a point in the plane is a combination of a horizontal and a vertical motion. So we say that a point has 2 degrees of freedom. If a frame- work on n vertices has no edge then it has 2n degrees of freedom. Adding an edge to the framework reduces its degrees of freedom by at most 1. However, for n ≥ 2, there are always motions that can not be blocked by edges, they are congruent motions of the framework. We can count for these 3 independent motions: 2 for translations and 1 for rotation. Therefore, in an infinitesimally rigid framework without redundant edges one should expect that
1. the framework has 2n − 3 edges in total, and 2. in each subframework with n′ vertices, there are at most 2n′ − 3 edges.
The necessity of these conditions has been known since James Clerk Maxwell’s time [81]. Laman showed that, for generic frameworks, it is also sufficient. In dimension 3 and higher, up to now, no combinatorial characterization is known for generic rigidity. Nevertheless, for the special class of frameworks that describe molecular structures, a long-conjectured combinatorial characterization, known un- der the name of Molecular Conjecture (by Tay and Whiteley 1984), has been proven to be true. Softwares based on this characterization such as FIRST, FRODA,. . . ([67],[105]) had been developed even before the confirmation of its mathematical correctness. Let us move to other notions of rigidity. The first one concerns with the global uniqueness of frameworks, namely, if the set of distance constraints imposed by bars is sufficient to determine the framework up to congruence. As an exam- ple, let us consider the sensor network localization problem. In a sensor network, autonomous sensors collect information on enviroment condition such as tempera- tures, pressures, chemical agents,. . . and send data through the network to a center. Though not all sensors are equipped with GPS receivers, which are costly and bat- tery consuming, to localize directly their position, the distance between some pairs of sensors can be calculated (using radio signals for close sensors, for example). A
12 Chapter 1. Introduction
• •
• • • • △ •
△ △ ••
(a) (b)
Figure 1.4: (a) A wireless sensor network. Elements with GPS receivers are represented with triangles and the others with black circles. The arrow lines represent the communicability between sensors. (b) The corresponding bar-joint framework. well-designed sensor network should allow one to locate all sensors based on this information. Such a network can be converted to a bar-joint framework by putting edges between pairs of sensors whose distances are known as well as between all sensors possessing a GPS receiver (Figure 1.4). It is evident that the first condition for the localizability is that the location of each sensor is uniquely determined by the available information on the distances. It is equivalent to the global rigidity of the corresponding framework. Although global rigidity looks quite different from local rigidity, they are actu- ally tightly related. It is easy to see that global rigidity implies local rigidity. In the other direction, it is shown that a planar generic framework is globally rigid if and only if one must remove at least 3 vertices to disconnect it and it remains locally rigid after the removal of any edge. In practice, it is also important that the location of the sensors can be computed efficiently. To this end, semidefinite programming-based algorithms are proposed. However, even when the location of the sensors is unique in the considered space (plane or 3-space), these algorithms may find a configuration in a higher dimen- sion. To prevent this undesirable outcome, universal rigidity, a stronger notion than global rigidity is of order. A framework is universally rigid if it is uniquely determined up to congruence in any space. Characterizing combinatorially univer- sal rigidity of generic frameworks seems more difficult than global rigidity, partly due to the fact that it does not depend uniquely on the underlying graphs, even when the vertices are in generic position (Figure 1.5). As a matter of fact, even in
13 1.1. Frameworks and rigidity
• •
• • • • •• ••
Figure 1.5: The planar generic framework on the left is universally rigid while the one on the right is not. dimension 1, no combinatorial characterization of universal rigidity is known. Driven by practical applications, many types of frameworks are extensively studied. Rigidity properties in these models are defined in the same manner as for bar-joint frameworks: local/infinitesimal rigidity for the uniqueness under continu- ous/infinitesimal motions, global rigidity for the uniqueness in the same dimension and universal rigidity for the uniqueness in all dimensions. We brief here some models studied in our thesis.
1. Direction-length frameworks: This is an extended model of bar-joint frame- works where beside distance constraints we have also direction constraints between vertices.
2. Body-bar frameworks and body-bar frameworks with boundaries. In a body- bar framework, bars link solid bodies and in a body-bar framework with boundaries, some bodies are linked to a fixed external environment by bars or they are simply pinned down. These models can be converted to bar-joint model but are interesting in their own right as combinatorial characterizations for rigidity in these models are obtained in all dimensions.
3. Body-length-direction frameworks: These are extended models of body-bar frameworks where we have both distance and direction constraints between bodies. We also consider different types of bodies which allow different types of motions.
The study of these frameworks may have applications in Computer-Aided-Design or in sensor network localization.
4. Tensegrity frameworks are a rather different model. In these frameworks we have three types of edges corresponding to three types of constraints: bars which fix distances between pairs of vertices, cables which prevent distances
14 Chapter 1. Introduction
from increasing and struts which prevent distances from decreasing. The consideration of these frameworks arose naturally in architectural engineering since materials which resist well compression are not strong in withstanding tension and vice versa. Using cable and strut-like elements helps reduce the cost and mass of the structure. Recent researches also propose considering cell structure as a tensegrity structure [56].
1.2 Our contributions and organization of the thesis
Chapter 2 reviews basic notions and facts used in our thesis, especially on graphs and digraphs, combinatorial optimization tools (matroid theory, submodular func- tion,. . . ) as well as algebra and linear algebra. Chapter 3 gives a basic formal introduction to the rigidity theory of bar-joint frameworks. The main results of this thesis appear from Chapter 4 to Chapter 8. Below, we give a summary of these chapters.
Chapter 4: Matroid approach. This chapter presents our early approach to the problem of characterizing local/infinitesimal rigidity in generic frameworks. View- ing this problem as characterizing the generic rigidity matroid, we study abstract rigidity matroids, a generalization of the generic rigidity matroid. We solve an open question by Graver, Servatius and Servatius [47] on the characterization of abstract rigidity matroids in dimension 2 and extend the result to all dimensions. We also introduce the notion of 1-extendable abstract rigidity matroids and inves- tigate the relation between these matroids and the generic rigidity matroids. These results compose an article appeared in SIDMA [85]. Furthermore, inspired by the counting condition by Laman, we propose the study of intersecting submodular functions that induce abstract rigidity matroids and obtain a characterization for these functions.
The core of our thesis lies in constructive approaches to the problem of rigidity. By “constructive approach” we first mean an approach that focuses on inductive constructions of rigid frameworks or graphs and the effect of extension operations on frameworks or graphs. This approach has proved its power particularly in characterizing infinitesimal/local rigidity (see, e.g., [101, 99, 98, 94, 72]). It is also sucessfully employed in characterizing global rigidity in [58, 24]. Second, an alternative to inductive construction is to find a decomposition for underlying graph
15 1.2. Our contributions and organization of the thesis of (minimally) rigid frameworks. Such a decomposition can be used to construct an explicit rigid framework. This approach often results in a simpler proof for the desired characterization. (See [100, 107, 60] for examples.)
Chapter 5: Inductive construction and decomposition of graphs. First, in Section 5.2, we provide an inductive construction and a decomposition for graded sparse graphs. These graphs arise when one considers frameworks with mixed constraints, then different types of edges are subject to different sparsity conditions. We also study the graded sparse matroids determined by these graphs, obtain the rank formula as well as a decomposition of these matroids. The result in this section is from a joint work with Bill Jackson [66]. Section 5.3 considers another notion of sparse graphs: (b, l)-sparse graphs, where the sparsity of subgraphs depend on their vertices. The motivation for this notion comes from the consideration of frameworks with bodies of different dimensions. We give an inductive construction for these graphs and also a charac- terization of these graphs as resulting graphs in a pebble game. Our results on these different types of sparse graphs generalize the classic result of Nash-Williams [83] on the decomposition of graphs into edge-disjoint spanning trees. The dual result of this is about packing of edge-disjoint spanning trees by Tutte [104] and Nash-Williams [82]. Their results can be obtained easily from the directed counterpart on packing of arc-disjoint arborescences by Edmonds [31] via an orientation result by Frank. Our third contribution in Chapter 5 is a generalization of the result of Edmonds to packing of arborescences whose roots are constrained to some matroidal condi- tion (Section 5.4). Using a general orientation result by Frank [38], we obtain a short proof for a result of Katoh and Tanigawa which generalizes Tutte and Nash- Williams’ result for characterizing rigidity of frameworks with boundaries. This is the result of a joint work with Olivier Durand de Gevigney and Zolt´an Szigeti which appeared in SIDMA [28].
Chapter 6: Infinitesimal rigidity. This chapter focuses on the infinitesimal rigidity of several types of frameworks with mixed constraints. In Section 6.3, we obtain combinatorial characterizations of infinitesimal rigidity in generic body- length-direction frameworks, using the decomposition for graded-sparse graphs in Chapter 5. This part is also from the previously mentioned joint work with Bill Jackson. Section 6.4 discusses extension operations for direction-length frameworks in general dimension. We extend the definition of 0-extension and 1-extension in R2
16 Chapter 1. Introduction to d-dimension case and investigate the effect of these extension operations on the infinitesimal rigidity and boundedness of generic frameworks. Chapter 7: Global rigidity of direction-length frameworks. We extend a result of Jackson and Jord´an on the global rigidity preservingness of 1-extensions on generic direction-length frameworks in R2 to all dimensions. This chapter consists of an article appeared in IJCGA [86]. Chapter 8: Universal rigidity. While local/infinitesimal rigidity and global rigidity are well studied with abundant results, especially for generic frameworks in dimension 1 and 2, little is known about universal rigidity even in dimension 1. We offer the study of universal rigidity in two directions. First, in Section 8.2, we explore universal rigidity on the line. We obtain a complete characterization for universal rigidity of complete bipartite frameworks on the line and show that the only generically universally rigid bipartite graph in d R is the single edge K1,1 for all d ≥ 1. Many open questions and conjectures inspired by our result are discussed. This section consists of a joint article with Tibor Jord´an [70]. The second direction is to relax the condition on the genericity of frameworks. In Section 8.3, we strengthen earlier results on sufficient condition for universal rigidity of bar-joint frameworks [8] and tensegrity frameworks [19] to allow config- urations in non-general position. This is the result from a joint work with Abdo Alfakih [6].
17 1.2. Our contributions and organization of the thesis
18 Chapter 2
Preliminaries
Contents 2.1 Basic notions ...... 20 2.2 Graphs and digraphs ...... 20 2.3 Matroid theory ...... 23 2.4 Submodular functions ...... 26 2.5 Algebra and linear algebra ...... 27 2.6 Congruent motions and motions of rigid bodies . . . . 30
19 2.1. Basic notions
2.1 Basic notions
We will use Z and Z+ to denote the set of integers and non negative integers, respectively. The set of rational numbers and real numbers are denoted by Q and R respectively. Rn denotes the n-dimensional Euclidean space. 0 denotes a zero matrix of appropriate dimensions or the origin of an Euclidean space. The n Euclidean norm of a vector/point x = (x1, . . . , xn) ∈ R is denote by kxk, that is
2 2 2 kxk = x1 + ··· + xn.
n The number of 2-element subsets of an n-element set is denoted by 2 . For a finite set S the number of elements of S is denoted by |S|. The collection of all the subsets of S is denoted by 2S. If a set A is a subset of a set B then we write A ⊆ B. We also write A ⊂ B when A ⊆ B and A =6 B and say that A is a proper subset of B.A multiset is a generalization of the notion of set, where each member may appear more than once. For two sets A, B, let A \ B or A − B denote the set of all elements belonging to A but not to B. A ∪ B and A ∩ B denote the union and intersection of A and B respectively. We also write A − x for A \{x} and A + x or A ∪ x for A ∪ {x}.
A partition of a non-empty set S is a collection P = {P1,...,Pn} where
P1,...,Pn are non-empty pairwise disjoint subsets of S such that P1 ∪· · ·∪Pn = S. Let f : X → Y be a map, A ⊆ X, and B ⊆ Y . The image of A by f is f(A) = {f(x): x ∈ A} and the pre-image of B by f is f −1(B) = {x ∈ A : f(x) ∈ B}. n m The Jacobian matrix of a differentiable map f : R → R , f(x1, . . . , xn) =
(f1(x1, . . . , xn), . . . , fm(x1, . . . , xn)), at a point p is the m × n matrix
∂f ∂f 1 ... 1 ∂x1 ∂xn . . . df|p = . .. . ∂f ∂f m m ... ∂x1 ∂xn 2.2 Graphs and digraphs
A graph (or undirected graph) is a pair G = (V,E) where V is a non-empty finite set and E consists of 2-element multisubsets of V . The elements of V are called vertices or nodes of G and the elements of E are called the edges of G. Sometimes, to be more precise, we use V (G) and E(G) to refer to the vertex set and the
20 Chapter 2. Preliminaries edge set of G. An edge {u, v} of G is often denoted by uv and u, v are called the extremities or ends of the edge uv. We say that the edge uv is incident to u and v. If u = v then the edge uv is called a loop. A graph is often depicted by a set of dots for the vertices and line segments or curves for the edges. A graph without loops and multiple edges is called a simple graph. The degree of a verter v in a graph G, denoted by d(v), is the number of non-loop edges incident to v plus twice the number of loops at v. Two simple graphs G = (V,E) and G′ = (V ′,E′) are said to be isomorphic if there exists a bijection φ : V → V ′ such that for every u, v ∈ V , uv is an edge of G if and only if φ(u)φ(v) is an edge of G′. A complete graph is a simple graph where between each pair of vertices there is an edge. A complete graph on n vertices is often denoted by Kn. We also use K(V ) to denote the complete graph on a vertex set V .A bipartite graph is a graph where the vertex set can be partitioned into two sets X and Y such that there is no edge between two vertices in X and there is no edge between two vertices in Y . A complete bipartite graph is a bipartite graph with a vertex partition X,Y such that for each pair of vertices x ∈ X and y ∈ Y , xy is an edge. If |X| = m and |Y | = n, the complete bipartite graph with the vertex partition X,Y is denoted by Km,n. Let G = (V,E) be a graph and X,Y be subsets of V . Throughout this thesis, we will use E(X) to denote the set of edges induced by X in G, i.e., those edges of G with both extremities in X, and i(X) to denote |E(X)|. If X = {v}, i(X) = i(v) is simply the number of loops at v. We will use E(X,Y ) to denote the set of edges from X to Y , i.e., those edges of G with one extremity in X and the other in Y . We denote |E(X,Y )| by i(X,Y ). A subgraph of G = (V,E) is a graph H = (V ′,E′) with V ′ ⊆ V and E′ ⊆ E(V ′). It is called an induced subgraph of G if E′ = E(V ′). The vertex set of an edge set F is denoted by V (F ). A path from a vertex u to a vertex v in a graph G = (V,E) is a sequence u = v0, e1, v1, e2, . . . , vk−1, ek, vk = v of vertices v0, . . . , vk in V and edges e1, . . . , ek in E such that vi−1, vi are two ends of ei for i = 1, . . . , k. We often identify a path with its sequence of edges. The edge set may be empty if u = v. If there is a path from u to v in G then we say that v is reachable from u. A connected graph is a graph G = (V,E) whose vertex set V can not be parti- tioned into two nonempty subsets X and Y such that i(X,Y ) = 0. Equivalently, a connected graph is a graph G such that, for each pair of vertices u, v, there is a path from u to v in G.A connected component of a graph G is a maximal con-
21 2.2. Graphs and digraphs nected subgraph of G. Obviously, connected components of G are vertex-disjoint. An edge e of G is called a cut-edge if the deletion of e from G increases its number of connected component. A cycle is a connected graph in which each vertex has degree 2. If an edge of G is not a cut-edge then it must belong to a cycle in G.A forest is a graph that does not contain any cycles as its subgraphs. A tree is a connected forest. A subgraph H of G is said to be spanning if V (H) = V (G). A tree is a spanning tree of G if it is a spanning subgraph of G. Let k be a positive integer. A graph G is said to be k-vertex-connected, or simply k-connected, if |V (G)| > k and one has to remove at least k vertices to disconnect the graph. Note that 1-connected graphs are simply connected graphs. Let P be a partition of the vertex set V of G into non-empty subsets. We denote by EG(P) the set of all edges in E(X,Y ) for every X,Y ∈ P,X =6 Y . Edges in
EG(P) are called crossing edges of the partition P. We often denote |EG(P)| by eG(P). A directed graph (or a digraph) is a pair D = (V,A) where V is a finite set and A consists of ordered pairs of elements of V (where we can have pairs of the same element). The elements of V are called the vertices of D and the element of A are called the arcs of D. An arc (u, v) is often denoted by uv; u is called the tail and v is called the head of the arc uv. If u = v then the arc uv is called a loop. A digraph is depicted in the same way as an undirected graph except that we use an arrow to denote the direction from u to v of an arc uv. An arc uv from u to v is said to be an out-arc of u and an in-arc of v. Let D = (V,A) be a digraph and X a subset of V . An arc uv with u ∈ V \ X and v ∈ X is said to be an entering arc of X. An outgoing arc of X is an entering − arc of V \X. The set of all arcs of D entering X is denoted by RD(X), the number − of its elements is denoted by ρD(X). When X = {v}, we simply write RD(v) and + ρD(v). Concerning the outgoing arc of X we will use RD(X), δD(X) respectively. A dipath from a vertex u to a vertex v in a digraph D = (V,A) is a sequence u = v0, a1, v1, a2, . . . , vn−1, an, vn = v of vertices v0, . . . , vn in V and arcs a1, . . . , an in A such that vi−1 is the tail and vi is the head of the arc ai for i = 1, . . . , n. We often identify a dipath with the sequence of its arcs. If there exists a dipath from v to u in D then we say that u is reachable from v in D. We say that D is an r-arborescence if D is a directed tree, r is a vertex of D of in-degree 0 and all the other vertices of D are of in-degree 1. We note that an r-arborescence may consist of only the vertex r and no arc. Note also that
22 Chapter 2. Preliminaries an r-arborescence has a unique vertex of in-degree 0, namely r. We also use out- arborescence to refer to an arborescence when we want to distinguish it with an in-arborescence, which is a directed graph where one vertex has out-degree 0 and all the other vertices have out-degree 1. A sub-digraph H of D is called spanning if its vertex set V (H) coincides with V . A digraph D is strongly connected if for every two vertices u, v of D, u is reachable from v and vice versa. D is said to be k-connected if |V (D)| > k and one need to remove at least k vertices to make D non strongly connected.
2.3 Matroid theory
Definition 2.3.1 (Matroid). A matroid M is a pair (S, I) of a finite set S and a collection I of subsets of S that satisfies the following independence axioms.
(I0) ∅ ∈ I. (I1) If I ∈ I and J ⊆ I then J ∈ I. (I2) If I,J ∈ I and |I| > |J| then there exists x ∈ I \ J such that J ∪ x ∈ I.
Matroid is a structure that generalizes the notion of linear independence. If S is a finite set of vectors in a vector space, the collection I of linearly independent subsets of S verifies the three properties above. We call this matroid the linear matroid defined by S. Matroid theory draws also motivations from graph theory. Given a graph G = (V,E), the edge sets of all subgraphs of G that are forests form the independent sets of a matroid on E. We call this matroid the graphic matroid of G. In fact, this matroid is also a linear matroid. The set S is called the ground set of the matroid M and the sets in I are called the independent sets. Sets that are not independent are said to be dependent. Minimal dependent sets are called circuits. If {x} is a circuit then x is called a loop. Two elements x, y of a matroid M are said to be parallel if {x, y} is a circuit of M. A base of a matroid M is a maximal inclusionwise independent set. The col- lection B of bases of a matroid M satisfies the following properties.
(B0) B is not empty.
(B1) If B1 and B2 are in B then |B1| = |B2|.
23 2.3. Matroid theory
(B2) (Exchange axiom) If B1,B2 ∈ B and x ∈ B1 \ B2 then there exists
y ∈ B2 \ B1 such that B1 − x + y ∈ B.
These axioms on bases can be used to define a matroid. In fact, if a collection B of subsets of a finite set S verifies the axioms (B0), (B1), (B2) then it determines uniquely a matroid by defining the independent sets as all the subsets of bases. Furthermore, if we set B∗ = {S \ B : B ∈ B} then B∗ also satisfies (B0), (B1), (B2) and therefore determines a matroid on S. We call this matroid the dual matroid of M. Bases and circuits of the dual matroid of M are called cobases and cocircuits of M. The complement of a cocircuit is called a hyperplane of M. Let M = (S, I) be a matroid. The rank of a set X ⊆ S is the maximum cardinality of an independent subset of X:
rM(X) = max{|I| : I ⊆ X,I ∈ I}.
When the matroid is clear from the context, we may write r(X) instead of rM(X). The rank of a matroid satisfies the following properties.
(R1) 0 ≤ r(X) ≤ |X| for every X ⊆ S. (R2) If X ⊆ Y then r(X) ≤ r(Y ) for every X,Y ⊆ S. (R3) (Submodularity) If X,Y ⊆ S then r(X) + r(Y ) ≥ r(X ∪ Y ) + r(X ∩ Y ).
A matroid is also determined uniquely by its rank function, i.e, if r : 2S → Z+ satisfies the three axioms (R1), (R2) and (R3) then it defines a matroid on S: the independent sets are subsets X of S with r(X) = |X|.
The closure operator clM(·) of a matroid M on S is an operator on the collection of subsets of S defined by
clM(X) = {x ∈ S : r(X ∪ x) = r(X)}. for every X ⊆ S. We may write cl(X) when M is clear from the context. Given a matroid M = (S, I) on a finite set S and a subset S′ of S, we can obtain a matroid on S′ by defining
I′ = {I ∈ I : I ⊆ S′}.
24 Chapter 2. Preliminaries
It is easy to see that I′ = {I ∩ S′ : I ∈ I}. The fact that I′ verifies (I0), (I1) and (I2) follows from that of I. Hence, I′ forms the independent sets of a matroid on S′. We call this matroid M′ = (S, I′) the restriction of M on S′. A free matroid on a finite set S is the matroid on S whose independent sets are all the subsets of S.
Count matroids and sparse graphs
Let us consider a class of matroids that plays an important role in combinarial rigidity theory: count matroids.
Let G = (V,E) be a graph and b : V → Z+. Here we use the bold symbol to emphasize the fact that b is a map. For the image of v ∈ V and X ⊆ V we write b(v), b(X) respectively. Let bmin denote min{b(v): v ∈ V }. Suppose that l is an integer with 0 ≤ l < 2bmin for every uv ∈ E. Then it is not difficult to show that the collection
I = {F ⊆ E : iF (X) ≤ b(X) − l for every X ⊆ V with iF (X) > 0} where iF (X) denotes the number of edges in F with both extremities belonging to X, forms the independent sets of a matroid on E (see, e.g., [39, Theorem 13.5.1] for a proof). Note that, by our assumption 0 ≤ l < 2bmin, the condition for the independence of an edge set F implies that if |X| ≥ 2 then iF (X) ≤ b(X) − l. The matroid (E, I) is called a (b, l)-count matroid on G. If b(v) = 1 for all v ∈ V and l = 1, the (b, l)-count matroid on G coincides with the graphic matroid on G. When b(v) = 2 for all v ∈ V and l = 3, the (b, l)-count matroid coincides with the 2-dimensional generic rigidity matroid (more details in Chapter 3). A graph G with the edge set independent in a (b, l)-count matroid on G is called a (b, l)-sparse graph.A(b, l)-tight graph is a (b, l)-sparse graph G = (V,E) with |E| = b(V ) − l. When all b(v) take the same value m, we also refer to these graphs as (m, l)-sparse graphs and (m, l)-tight graphs, respectively. In some context, to emphasize the sparseness, we also say tight sparse graphs. Familiar examples are (1, 1)-sparse graphs which are actually forests, and (1, 1)-tight graphs which are trees.
Matroid union
One way to get a matroid from other matroids is taking their unions. Let M1 =
(S1, I1),..., Mk = (Sk, Ik) be k matroids. The matroid union of M1,..., Mk is
25 2.4. Submodular functions
the matroid M = (S, I) where S = S1 ∪ · · · ∪ Sk and
I = {I1 ∪ · · · ∪ Ik : I1 ∈ I1,...,Ik ∈ Ik}.
If Si =6 S, we can extend the matroid Mi to a matroid on S by defining all elements of S \ Si to be loops. We identify this new matroid with Mi. Therefore, for the sake of convenience, we may suppose that Mi are matroids on the same ground set S. Many fundamental results become clear when viewed under the matroid union angle. For instance, the result of Nash-Williams that (k, k)-sparse graphs are the union of k edge-disjoint forests can be stated as every (k, k)-sparse matroid is the union of k (1, 1)-sparse matroids (more details in Chapter 5). The rank function for the matroid union is determined as in the following result. (For a proof, see e.g. [87, Theorem 12.3.1 ].)
Theorem 2.3.1. Let M be the matroid union of M1,..., Mk with rank functions rM1 , . . . , rMk . Then the rank function of M is given by
rM(X) = min{rM1 (Y ) + ··· + rMk (Y ) + |X \ Y | : Y ⊆ X} (2.3.1) for every X ⊆ S.
2.4 Submodular functions
Definition 2.4.1 (Submodular function). Let S be a finite set. An integer-valued function f : 2S → Z is said to be submodular if it satisfies
f(X) + f(Y ) ≥ f(X ∪ Y ) + f(X ∩ Y ) (2.4.1) for every subset X,Y of S.
If f : 2S → Z satisfies inequality (2.4.1) for every X,Y ⊆ S with X ∩ Y =6 ∅ then it is called an intersecting submodular function. A function g : 2S → Z is a supermodular function (intersecting supermodular function, resp.) if −g is submodular (intersecting submodular, resp.). Submodular functions (and hence supermodular functions) play an important role in combinatorial optimization because of its omnipresence. For example in a graph G, ρ(X) is a submodular function, i(X) is a supermodular function, for X ⊆ V (G), and m|V (F )|−l is an intersecting submodular function, for F ⊆ E(G), where 1 ≤ m < 2l. Moreover, submodular functions are of special interest in combinatorial optimization as they can be minimized in polynomial time.
26 Chapter 2. Preliminaries
Theorem 2.4.1 (Iwata, Fleischer and Fujishige [57], Schrijver [92]). Given a sub- modular function f : 2S → Z, a set U ⊆ S that minimizes f(U) can be found in polynomial time.
Submodular functions are closely related to matroids. As mentioned above, the rank of a matroid is a submodular function. Conversely, given a nondecreasing intersecting submodular function f : 2S → Z, let us consider the collection
I(f) = {I ⊆ S : |J| ≤ f(J) for every J ⊆ I,J =6 ∅}.
Theorem 2.4.2. I(f) forms the independent sets of a matroid M(f) on S. The rank function of M(f) is determined by
t
rM(f)(X) = min{|X0| + f(Xi): {X0,X1,...,Xt} partitions X} (2.4.2) i=1 X for X ⊆ S.
(For a proof, see [39], for example.) The circuits of this matroid are minimal non empty sets C such that f(C) < |C| and f(C) ≥ f(C − x) ≥ |C| − 1 for every x in C. Therefore, we have the following.
Proposition 2.4.3. If C is a circuit of the matroid M(f) induced by a nonde- creasing intersecting submodular function f then f(C) = f(C − x) = |C| − 1.
2.5 Algebra and linear algebra
Let K be a subfield of the complex field C. By K[X1,...,Xn] we denote the ring of all n-variable polynomials f[X1,...,Xn] with coefficients in K.
The extension field of a field K ⊆ C by p1, . . . , pn ∈ C, denoted by K(p1, . . . , pn), is the smallest subfield of C that contains p1, . . . , pn. Equivalently, K(p1, . . . , pn) =
{f(p1, . . . , pn)/g(p1, . . . , pn): f, g ∈ K[X1,...,Xn], g(p1, . . . , pn) =6 0}. The algebraic closure of a field K ⊆ C is the smallest field K ⊆ C such that every polynomial f(X1) ∈ K[X1] has a root in K.
A set {p1, . . . , pn} ⊂ C is said to be algebraically independent over K if there does not exist a polynominal f(X1,...,Xn) ∈ K[X1,...,Xn], non identical to zero, such that f(p1, . . . , pn) = 0. It turns out that this algebraic independence is in fact matroidal.
27 2.5. Algebra and linear algebra
Theorem 2.5.1 ([87] Theorem 6.7.1). Let L be an extension field of a field K and S a finite subset of L. Then the collection I of subsets of S that are algebraically independent over K is the set of the independent sets of a matroid on S.
For an m × n matrix M we often use Mij to denote the (i, j)-entry of M. The T T transpose of M is the n × m matrix denoted by M such that Mij = Mji.
We use In to denote the identity matrix of order n, that is the n × n matrix with all (i, i)-entries being 1 and all the other entries being 0. When the dimension n is clear from the context we simply write I for In. An n × n real matrix M is T T said to be orthogonal if MM = M M = In. Sometimes we deal with matrices M whose entries are functions of a parameter d t. We denote such a matrix by M(t). The derivation M(t) is simply the matrix dt d whose (i, j)-entries are M(t) . It is routine to verify that dt ij d d d M(t)N(t) = M(t) N(t) + M(t) N(t). dt dt dt The nullspace Ker M of a matrix M with n columns is the linear space of all vector x ∈ Rn such that Mx = 0. For an n×n matrix M, let trace (M) denote the sum of all the diagonal entries of M. Let A be an n × m matrix and B an m × n matrix, then it is easy to verify that
trace (AB) = trace (BA).
An n × n-matrix A = (aij) is symmetric if aij = aji for all 1 ≤ i, j ≤ n. The set of all real n × n matrices is denoted by Sn. Then Sn is a vector space and we can define an inner product h· , ·i in Sn by
hA,Bi = trace (AB), for A, B ∈ Sn. Let M be an n × n real matrix. A scalar λ is said to be an eigenvalue of M if there exists x ∈ Rn, x =6 0 such that Mx = λx. A symmetric n × n matrix A is said to be positive definite if for every non-zero x ∈ Rn we have xT Ax > 0. A symmetric n×n matrix A is positive semidefinite (or PSD for short) if for every x ∈ Rn the inequality xT Ax ≥ 0 holds. PSD matrices are used to characterize the universal rigidity of frameworks (see Chapter 8). A principal minor of an n × n matrix A is the determinant of a square submatrix of A with rows and columns indexed by the same subset X of {1, . . . , n}. If X = {1, . . . , k} for some k ≤ n then the principal minor is called a leading principal minor. The following lemma summarizes equivalent statements about the positive semidefiniteness of a matrix.
28 Chapter 2. Preliminaries
Lemma 2.5.2. Let A ∈ Sn. The following statements are equivalent.
1. A is a PSD matrix.
2. All eigenvalues of A are non-negative.
3. All principal minors of A are non-negative.
4. All leading principal minors of A are non-negative.
5. A = XXT for some n × n matrix X.
A consequence of Lemma 2.5.2 is the following.
Lemma 2.5.3.
1. If A is a PSD matrix in Sn then trace (A) ≥ 0 and trace (A) = 0 if and only if A = 0.
2. Let A, B be two PSD matrices in Sn, then hA,Bi ≥ 0 and hA,Bi = 0 if and only if AB = 0.
Proof. 1. By Lemma 2.5.2 we may assume that A = XXT for some n × n matrix
X = (xij). Then
T 2 trace (A) = Xij(X )ij = xij ≥ 0. i,j i,j X X
Moreover, trace (A) = 0 if and only if xij = 0 for all i, j, which means that X = 0 and hence A = 0.
2. We may suppose that A = XXT and B = YY T for some n × n matrices X,Y . Then
hA,Bi = trace (AB) = trace (XXT YY T ) = trace (Y T XXT Y ) = trace ((XT Y )T (XT Y ) ≥ 0.
Moreover, trace (AB) = 0 if and only if the trace of the PSD matrix (XT Y )T (XT Y ) is 0, which implies XT Y = 0 by the first statement and hence AB = 0 holds.
A n × n matrix A = (aij) is skew-symmetric if aij = −aji for all 1 ≤ i, j ≤ n; in particular, aii = 0 for all 1 ≤ i ≤ n. Skew-symmetric matrices will be used to describe infinitesimal rotations of rigid bodies.
29 2.6. Congruent motions and motions of rigid bodies
d Let x, y be two vectors in R . The join of x = (x1, . . . , xd) and y = (y1, . . . , yd) d is the 2 -dimensional vector (1,2) (1,3) (i,j)(d−1,d)
x1 x2 x1 x3 xi xj xd−1 xd x ∨ y = , ,..., ,..., y1 y2 y1 y3 yi yj yd−1 yd !
d Let A = (a ij) be a d × d skew-symmetric matrix. We define a 2 -dimensional vector w associated with A by w = (a12, a13, . . . , a(d−1)d), i.e., by putting aij con- secutively in the lexicographic order. A related notion to join in R3 is the cross product
x x x x x x x × y = 2 3 , 3 1 , 1 2 y2 y3 y3 y1 y1 y2 !
which differs from x ∨ y in the order and a possible −1 scaling of entries. The following equality relates inner product and join.
hAy , xi = hw , x ∨ yi. (2.5.1)
Indeed,
hAy , xi = aijyjxi i,j X = aijyjxi + aijyjxi (since aii = 0 for every i) i 2.6 Congruent motions and motions of rigid bod- ies A congruence of Rd is a map h : Rd → Rd such that kh(p) − h(q)k = kp − qk for all p, q ∈ Rd. (2.6.1) The following result is fundamental, a proof can be found in [77, 22]. 30 Chapter 2. Preliminaries Proposition 2.6.1. A map h : Rd → Rd is a congruence if and only if there exists an orthogonal d × d matrix M such that h(p) = Mp + h(0) for all p ∈ Rd. A congruent motion of Rd is a map P : [0, 1] × Rd → Rd such that P (t, p) is continuously differentiable in p and t and for each t ∈ [0, 1], the map P (t, ·): Rd → Rd is a congruence, i.e., P (t, p) = M(t)p + P (t, 0) for all t ∈ [0, 1] and p ∈ Rd, (2.6.2) where M(t) is a d × d orthogonal matrix. d Let A(t) = M(t), differentiating both sides of the equality M(t)M T (t) = I dt we have, M(t)A(t)T + A(t)M(t)T = 0. At t = 0, M(0) = I, so we obtain A(0)T + A(0) = 0, i.e., A = A(0) is a skew- symmetric matrix. By 2.6.2, the infinitesimal motion of p induced by the congruent motion P , i.e., the instantaneous velocity of p at instant t = 0, is given by d P (t, p)| = Ap + t. dt t=0 d Here t = P (t, 0)| which corresponds to the instantaneous translation, while dt t=0 A corresponds to the instantaneous rotation. Inversely, suppose that A is a d × d skew-symmetric matrix and t ∈ Rd. Let t2 tn M(t) = I + tA + A2 + ··· + An + · · · ≡ etA. 2! n! Then it is easy to show that this series converges for every t ∈ R, so M(t) is well-defined. Moreover, M(t) is infinitely differentiable and d M(t)| = A. dt t=0 T T On the other hand, M(t)M(t)T = etAetA = etA+tA = I, since A is skew- symmetric. Hence M(t) is orthogonal and P (t, p) = M(t)p + tt for all t ∈ [0, 1] and p ∈ Rd, is a congruent motion whose infinitesimal motion is Ap + t for all p ∈ Rd. 31 2.6. Congruent motions and motions of rigid bodies Therefore, an infinitesimal motion of a congruent motion (or infinitesimal con- gruence for short) of Rd can be described by a pair (A, t) of a skew-symmetric d×d matrix A and a vector t ∈ Rd. In particular, the space of infinitesimal congruences d d+1 of R is of dimension 2 . A rigid body is an idealization of a solid body in physical world. In a rigid body, the distance between any two points is constant. When there is no specification, a rigid body B in Rd, or simply a body, is understood to be of full dimension, i.e, the set of points of B affinely spans Rd A motion of a rigid body B ⊂ Rd is a map P : [0, 1] × B → Rd such that, for every p ∈ B, the map P (·, p) is continuous and kP (t, p) − P (t, q)k = kP (0, p) − P (0, q)k, for every t ∈ [0, 1], (2.6.3) that is, the motion preserves the distance between points in the rigid body B, or equivalently speaking, P (t, ·) is an isometry of B for every t ∈ [0, 1]. In fact, P (t, p) is the position of a point p in B at time t. These isometries P (t, ·) can be extended uniquely to congruences of Rd [47, 22] P (t, p) = M(t)p + P (t, 0). The above discussion shows that the infinitesimal motion of a rigid body in Rd can also be described by a pair (A, t) of a skew-symmetric matrix A and a vector t ∈ Rd. 32 Chapter 3 Rigidity theory Contents 3.1 Introduction ...... 34 3.2 Various types of rigidity ...... 35 3.2.1 Local rigidity ...... 35 3.2.2 Infinitesimal rigidity and rigidity matrix ...... 36 3.2.3 Local rigidity versus infinitesimal rigidity ...... 38 3.2.4 Static rigidity ...... 40 3.2.5 Infinitesimal rigidity versus static rigidity ...... 41 3.2.6 Global rigidity ...... 41 3.2.7 Universal rigidity ...... 42 3.2.8 Stress matrices ...... 42 3.3 Combinatorial rigidity ...... 44 33 3.1. Introduction 3.1 Introduction The theory of rigidity is characterized by its diversity from many aspects: moti- vations and applications, techniques, and, especially, models. It is thus difficult to cover the basic concepts in all prevailing models in this modest introduction into rigidity theory. In this chapter, we content ourselves with a description of the theory of rigidity for the bar-joint model. The first reason is that this model is easy to describe. Secondly, many other models can be converted to this model. Thirdly, ideas and techniques from the study of this model can be applied to other models as well. Lastly, although being quite simple to describe, the problems of characterizing the rigidity in the bar-joint model remains among the most difficult problems in rigidity theory. Therefore, bar-joint model serves as a base model to understand basic concepts, techniques as well as challenges in rigidity theory. A d-dimensional bar-joint framework is a pair (G, p) of a simple graph G = (V,E) and an embedding p which maps each vertex v ∈ V to a point p(v) in Rd. The vertices and the edges of G model the joints and the bars of a bar-and-joint structure while p is the placement of the structure in the d-dimensional Euclidean space. The pair (G, p) is also called a d-dimensional bar-joint realization of G and sometimes we refer to p as a configuration of V in Rd. The affine dimension of a configuration p is the dimension of the affine space spanned by {p(v): v ∈ V }. We say that a configuration p or a framework (G, p) in Rd is of full dimension if the affine dimension of p is d. The embedding p can also be regarded as a point p ∈ Rd|V |. Two realizations (G, p) and (G, q) of G in Rd are said to be equivalent if kp(u) − p(v)k = kq(u) − q(v)k, for every edge uv in E, i.e., the length of the bars are the same in these two realizations. They are said to be congruent if kp(u) − p(v)k = kq(u) − q(v)k, for every pair of vertices u, v in V. Definition 3.1.1 (Rigidity map). The d-dimensional rigidity map of a graph G is d|V | |E| the map fG : R → R defined by uv∈E 2 T fG(p) = (..., kp(u) − p(v)k ,... ) Note that two frameworks (G, p) and (G, q) are equivalent if and only if fG(p) = fG(q). They are congruent if fK(V )(p) = fK(V )(q). 34 Chapter 3. Rigidity theory 3.2 Various types of rigidity This section introduces basic concepts about various types of rigidity and discusses the relation between these types. 3.2.1 Local rigidity The most visual way to talk about local rigidity is to use continuous deformations as in Chapter 1. We formalize this definition as follows. A continuous motion from a framework (G, p) to a framework (G, q) in dimension d is a map P : [0, 1] × V → Rd such that 1. P (0, v) = p(v) for all v ∈ V , 2. P (1, v) = q(v) for all v ∈ V , 3. P (·, v) is continuous for all v ∈ V , 4. kP (t, u) − P (t, v)k = kp(u) − p(v)k for all uv ∈ E(G) and t ∈ [0, 1]. We say that P is a smooth motion if P is a continuous motion and P (·, v) is infinitely differentiable for each v ∈ V . Gluck [43] showed that the following three definitions of local rigidity are equiv- alent. Definition 3.2.1 (Local rigidity–continuous). A framework (G, p) in Rd is said to be locally rigid if every continuous motion of (G, p) in Rd results in a framework (G, q) that is congruent to (G, p). Definition 3.2.2 (Local rigidity–topological). A framework (G, p) in Rd is locally nd rigid if there exists a neighborhood Np ⊂ R of p such that for every equivalent realization (G, q) with q ∈ Np we must have that (G, q) is congruent to (G, p). Definition 3.2.3 (Local rigidity–analytic). A framework (G, p) in Rd is locally rigid if every smooth motion of (G, p) in Rd results in a framework (G, q) that is congruent to (G, p). The topological definition above gave rise to the term “local rigidity”. Definition 3.2.4 (Flexible). A framework is called flexible in Rd if it is not locally rigid in Rd. 35 3.2. Various types of rigidity Using the rigidity map, we can also restate the topological definition of local rigidity as follows. A d-dimensional framework (G, p) is locally rigid if for a small enough −1 −1 neighborhood Np of p, fG (fG(p)) ∩ Np = fK(V )(fK(V )(p)) ∩ Np. 3.2.2 Infinitesimal rigidity and rigidity matrix The problem of determining the rigidity of a framework becomes more tractable if we linearize the problem. Suppose that P is a smooth motion of (G, p). Then kP (t, u) − P (t, v)k2 = kp(u) − p(v)k2 for all uv ∈ E(G) and t ∈ [0, 1]. ′ Differentiating this equation at t = 0 and setting µ(u) = P (t, u)|t=0, µ(v) = ′ P (t, v)|t=0, we have (p(u) − p(v))(µ(u) − µ(v)) = 0, for all uv ∈ E(G). Here, µ(u) can be regarded as the instantaneous velocity of the joint u at time t = 0. This motivates us to define the infinitesimal motions of a framework as follows. Definition 3.2.5 (Infinitesimal motions). An infinitesimal motion of a d-dimensional framework (G, p) is an assignment, to each vertex v ∈ V , of a vector µ(v) ∈ Rd such that (p(u) − p(v))(µ(u) − µ(v)) = 0, for all uv ∈ E(G). We can view an infinitesimal motion as a vector µ = (. . . , µ(u),... )T ∈ Rd|V |. Then it is not difficult to show that the infinitesimal motions of a framework (G, p) form a vector subspace of Rd|V |. Among the smooth motions of a framework, there are those arising from the congruent motions of space. The infinitesimal motions induced by these motions are called trivial infinitesimal motions. Again, these trivial infinitesimal motions form a vector subspace of the infinitesimal motion subspace. Definition 3.2.6 (Infinitesimal rigidity). A framework (G, p) is infinitesimally rigid if its only infinitesimal motions are trivial ones. Otherwise, it is said to be infinitesimally flexible. When the affine dimension of {p(v): v ∈ V } is d, the dimension of the trivial d+1 infinitesimal motion space is 2 (c.f. Section 2.6). Roughly speaking, there are 36 Chapter 3. Rigidity theory d d independent infinitesimal motions induced by translations and 2 independent infinitesimal motions induced by rotations. When this affine dimension is strictly less than d, i.e., when all the vertices lie in a hyperplane of Rd, all infinitesimal motions are trivial if and only if G is a complete graph on at most d vertices [47]. Given a framework (G, p) in Rd, the question whether (G, p) is infinitesimally rigid can be answered by considering the rank of the so-called rigidity matrix. Definition 3.2.7 (Rigidity matrix). The rigidity matrix R(G, p) of a d-dimensional framework (G, p) is a |E| × d|V | matrix whose rows are indexed by the edges of G and whose columns are indexed by the vertices of G such that • each edge indexes one row; • each vertex indexes d columns; • the submatrix indexed by an edge e = uv and the vertex u and v are p(u)−p(v) and p(v) − p(u) respectively; • elsewhere all entries are zero. That is, R(G, p) is written as u v . . . . e=uv ··· 0 ··· p(u) − p(v) ··· 0 ··· p(v) − p(u) ··· 0 ··· . . . . For example, the rigidity matrix of the framework in Figure 3.1 is 1 2 3 4 {1,2} 0 −1 0 1 0 0 0 0 {1,3} −1 −1 0 0 1 1 0 0 {1,4} −1 0 0 0 0 0 1 0 {2,3} 0 0 −1 0 1 0 0 0 {3,4} 0 0 0 0 0 1 0 −1 It is easy to see that the rigidity matrix is in fact half of the Jacobian of the rigidity map, i.e, 1 R(G, p) = df | . 2 G p It is also immediate from the definition that µ ∈ Rd|V | is an infinitesimal motion of (G, p) if and only if µ is in the nullspace of R(G, p). Moreover, we have the following. 37 3.2. Various types of rigidity p(2)=(0,1) p(3)=(1,1) • • • • p(1)=(0,0) p(4)=(1,0) Figure 3.1: A planar framework. Proposition 3.2.1 ([9]). Let (G, p) be a framework on n vertices in Rd with |V | ≥ 2. Then the rank of R(G, p) is at most S(n, d) where dn − d+1 , n ≥ d + 1; S(n, d) = 2 n , n ≤ d. ( 2 Moreover, a framework (G, p) on n vertices in Rd is infinitesimally rigid if and only if rank R(G, p) = S(n, d). It is worth remarking that S(n, d) is also the upper bound for the rank of the rigidity matrix of a complete framework on n vertices. 3.2.3 Local rigidity versus infinitesimal rigidity The infinitesimal rigidity is in fact a stronger property than local rigidity. Theorem 3.2.2. If a framework (G, p) is infinitesimally rigid then it is locally rigid. Several different proofs for this well-known fact can be found in [22] for ex- ample. Here we brief a proof idea by Alexandrov and Gluck [43]. If G has at most d vertices then by Proposition 3.2.1,(G, p) is infinitesimally rigid if and only if G is a complete graph, so obviously (G, p) is locally rigid. So suppose that G has at least d + 1 vertices. The fact that (G, p) is infinitesimally rigid im- plies that p is a regular point of fG and of fK(V ) as well. Therefore, for a small −1 enough neighborhood Np of p, fG (fG(p)) ∩ Np is a manifold whose co-dimension is rank R(G, p). But rank R(G, p) = rank R(K(V ), p) is also the co-dimension of −1 the manifold fK(V )(fK(V )(p))∩Np. This latter manifold is obviously a submanifold of the former one hence the equality between their co-dimension implies that they coincide. Therefore, (G, p) is locally rigid. The converse of Theorem 3.2.2 is not always true. Figure 3.2(a) illustrates a 2-dimensional framework which is locally rigid but not infinitesimally rigid.A 38 Chapter 3. Rigidity theory slightly different embedding of the same graph as in Figure 3.2(b) is however both locally rigid and infinitesimally rigid. In fact, the first embedding is “special” in some sense: the three vertices b, e, c are collinear. It is not the only “special embedding” that may cause the difference between local rigidity and infinitesimal rigidity. Figure 3.2(c) shows a 2-dimensional framework without three collinear vertices which is locally rigid but not infinitesimally rigid. Considering only generic embeddings helps us avoiding this unconvenient situation. a b c •a • •• •e f• d• d• •••e •• • • b c b c a d •e (a) (b) (c) Figure 3.2: (a) There is a nontrivial infinitesimal motion with µ(a) = µ(b) = µ(c) = µ(d) = 0 and µ(e) perpendicular to the line bec. (b) A generic embeding of the same graph as in (a). No non-trivial infinites- imal motion exists. (c) A non-generic framework where abcd is a rectangle and abe, bcf are equi- lateral right triangles. A non-trivial infinitesimal motion is depicted. Definition 3.2.8 (Linearly generic embedding). An embedding p : V → Rd is (linearly) generic if every submatrix of R(K(V ), p) attains the maximum rank over all d-dimensional embeddings. A framework (G, p) with p being a generic embedding is called a generic framework or generic realization of G. Theorem 3.2.3 (Asimow and Roth [9]). A generic framework is locally rigid if and only if it is infinitesimally rigid. Rigidity matroids An important property of generic embeddings is that for two arbitrary generic embeddings and a subset F of edges, the rank of the set of rows indexed by F in the two rigidity matrices are the same. Therefore, all generic embeddings define a unique linear matroid on the edge set E of G. We call this matroid the d- dimensional generic rigidity matroid of G. This matroid can be regarded as the 39 3.2. Various types of rigidity restriction on E(G) of the d-dimensional generic rigidity matroid of the complete graph K(V ). We denote the latter by Gd(n) where n = |V |. Note that in this matroid, edge sets that induce isomorphic subgraphs have the same rank, namely, ′ this matroid depends uniquely on n and d. Again, if n ≤ n , Gd(n) can be considered ′ as a restriction of Gd(n ), so sometimes we will use Gd to mean a large generic rigidity matroid that contains all the d-dimensional generic rigidity matroids considered in our context. 3.2.4 Static rigidity Let us consider rigidity from a structural engineering viewpoint. Suppose that at each joint v of a 3-dimensional bar-joint structure we apply a force f(v) such that the net force and the net moment (about three coordinate axes) on the whole structure is zero. Such a system F = (..., f(v),... ) of forces is called an equilibrium force. The equilibrium condition is equivalent to f(v) = 0 and f(v) × p(v) = 0, (3.2.1) v∈V v∈V X X which is a system of 6 linearly independent equations. The first three equations are for the three components of the net force and the last three equations are for the net moment about three coordinate axes. It follows that the space of all equilibrium forces of a 3-dimensional structure is a vector space of dimension 3|V | − 6. The structure is stable if for every equilibrium force, it can avoid the deformation by generating tensions and compressions on its bars – we say that the equilibrium force is resolved. The system (3.2.1) can be rewritten using join as (f(v), 0) ∨ (p(v), 1) = 0. v∈V X where (f(v), 0) and (p(v), 1) are 4-dimensional vectors obtained from f(v), p(v) by adding a 0 and a 1 entry respectively. For bar-joint frameworks in general dimensions d, we say that a system of forces F = (..., f(v),... ), f(v) ∈ Rd, exerted on the joints, is in equilibrium if (f(v), 0) ∨ (p(v), 1) = 0. v V X∈ (See [27] for more details.) Therefore, all equilibrium forces of a d-dimensional bar-joint framework of full d+1 dimension form a vector space which has dimension d|V | − 2 . 40 Chapter 3. Rigidity theory Definition 3.2.9 (Static rigidity). A framework (G, p) is statically rigid if for every equilibrium force F = (..., f(u),... ) there is an assignment of scalar ωuv to each edge uv ∈ E such that f(u) + ωuv(p(v) − p(u)) = 0, for every u ∈ V. (3.2.2) uv E X∈ Here, ωuv plays the role of the internal stress on the bar uv, which generates an internal force ωuv(p(u) − p(v)) at the joint u. Equation (3.2.2) means that the sum of the external force and the internal forces at each joint is zero. 3.2.5 Infinitesimal rigidity versus static rigidity Let F = (..., f(u),... ) be a system of forces in equilibrium on a framework (G, p) which is resolved by a stress ω = (. . . , ωuv,... ). Using the rigidity matrix, the system of linear equations (3.2.2) can be rewritten as F − ωR(G, p) = 0, which means that F belongs to the row space of R(G, p). Conversely, if F = aR(G, p), where a ∈ R|E(G)| is a vector in the row space of R(G, p) then it is an equilibrium force and obviously can be resolved with ω = a. Therefore, the framework (G, p) is statically rigid if and only if the row space of R(G, p) has d+1 dimension d|V | − 2 . But it is also the necessary and sufficient condition for (G, p) to be infinitesimally rigid. Hence we obtain the equivalent between the static rigidity and the infinitesimal rigidity of bar-joint frameworks. Theorem 3.2.4 (see, e.g., [47, 22]). A framework (G, p) is infinitesimally rigid if and only if it is statically rigid. 3.2.6 Global rigidity Definition 3.2.10 (Global rigidity). A framework (G, p) in Rd is globally rigid if for every framework (G, q) in Rd, (G, q) is equivalent to (G, p) implies that (G, q) is congruent to (G, p). From this definition, a globally rigid framework is clearly locally rigid. 41 3.2. Various types of rigidity 3.2.7 Universal rigidity So far, we have always considered frameworks and their motions in the same dimen- sion. As discussed in Chapter 1, the uniqueness of the location for a localization problem, which is equivalent to the global rigidity of the associated framework, does not imply that we can find this location efficiently. The universal rigidity of a framework is a stronger property that guarantees the efficiency of the exploited SDP method. Definition 3.2.11 (Universal rigidity). A framework (G, p) is said to be univer- sally rigid if and only if, for every framework (G, q) in any dimension, (G, q) is equivalent to (G, p) implies that (G, q) is congruent to (G, p). Universal rigidity as universal local rigidity It is clear from the definitions that universal rigidity is a stronger property than global rigidity which is again stronger than local rigidity. On the other hand, the following result of Bezdek and Connelly shows that, if (G, p) and (G, q) are two realizations of G in dimension d ≤ d′ respectively, then there is a smooth motion from (G, p) to (G, q) in dimension 2d′. (We regard (G, p) and (G, q) as embedded in dimension 2d′ by adding zeros to missing coordinates.) Lemma 3.2.5 (Leapfrog Lemma [14]). Suppose that p and q are two embeddings of G in Rm. Then the following P (t, v) is a smooth motion in R2m such that P (0, v) = p(v),P (1, v) = q(v) for v ∈ V and for 0 ≤ t ≤ 1, kP (t, u) − P (t, v)k is monotone for every u, v ∈ V : p(v) + q(v) p(v) − q(v) p(v) − q(v) P (t, v) = + (cos πt) , (sin πt) , for v ∈ V. 2 2 2 This result implies that universal rigidity can be regarded as local rigidity in a universal sense. 3.2.8 Stress matrices Definition 3.2.12 (Self-stress). A self-stress or an equilibrium stress of a frame- work (G, p) is an assignment of a scalar ωuv = ωvu to each edge uv ∈ E such that for each vertex u ∈ V the following equilibrium condition is satisfied. ωuv(p(u) − p(v)) = 0. (3.2.3) v∈V :uv∈E X 42 Chapter 3. Rigidity theory p(2)=(1,0) p(3)=(1,1) • 1 • −1 1 1 −1 • • p(1)=(0,0) 1 p(4)=(1,0) Figure 3.3: A framework with an equilibrium stress. From the statics viewpoint discussed above, a self-stress is a stress of the bar- joint structure which resolves the zero equilibrium force. It is also immediate from the definition that ω = (ωuv : uv ∈ E) is an equilibrium stress of (G, p) if and only if ωR(G, p) = 0. Definition 3.2.13 (Stress matrix). Let (G, p) be a framework and ω = (ωuv : uv ∈ E(G)) be an equilibrium stress of (G, p). The stress matrix of (G, p) associated with ω is the |V | × |V | symmetric matrix Ω with rows and columns indexed by vertices in V such that −ωuv if uv ∈ E, Ωuv = w∈V :uw∈E ωuw if u = v, 0 otherwise. P For example, the stress matrix associated with the stress in Figure 3.3 is 1 2 3 4 1 1 −1 1 −1 2 −1 1 −1 1 Ω = . 3 1 −1 1 −1 4 −1 1 −1 1 Global rigidity and universal rigidity via stress matrices It turns out that stress matrices encode rigidity properties of frameworks. Below we summarize some well-known results on the relations between stress matrices and rigidity of frameworks. These results are explained with more details in Chapter 8. First we need one more definition. Definition 3.2.14 (Algebraically generic embedding). An embedding p : V → Rd is algebraically generic if the set of all the coordinates of p is algebraically independent over the rationals. 43 3.3. Combinatorial rigidity Theorem 3.2.6 (Connelly [23], Gortler, Healy and Thurston [44]). An alge- braically generic framework (G, p) on n ≥ d + 2 vertices in Rd is globally rigid if and only if (G, p) possesses a stress matrix of rank n − d − 1. Theorem 3.2.7 (Connelly [21, 19], Alfakih [3], Gortler and Thurston [45]). An algebraically generic framework (G, p) on n ≥ d + 2 vertices in Rd is universally rigid if and only if (G, p) possesses a positive semidefinite stress matrix of rank n − d − 1. Furthermore, Alfakih and Ye [8] shows that the “if” part of Theorem 3.2.7 still holds if one relaxes the algebraic genericity condition on the embedding to a general position condition, which requires that, under the map p, every d + 1 vertices of V are mapped to affinely independent points in Rd. This result will be strengthened in Chapter 8. For the sake of convenience, in passages concerning different types of rigid- ity, we will use the term “generic” both for “linearly generic” if it is about lo- cal/infinitesimal rigidity and for “algebraically generic” if it is about global/universal rigidity. 3.3 Combinatorial rigidity One central problem of rigidity theory is that given a framework, decide whether it is rigid (locally, infinitesimally, globally or universally) or not. This problem for an arbitrary embedding is known to be untractable, in general [91], [1]. The equivalence between infinitesimal rigidity and local rigidity for generic frameworks as well as the characterization of globally and universally rigid algebraically generic frameworks through stress matrices suggest that this problem is more tractable for generic frameworks. On the other hand, it often turns out that if a rigidity property holds for a generic realization (G, p) of a graph G then it also holds for every generic realization (G, q) of G in the same dimension. If this happens, the rigidity property in question is said to be a generic property. More generally, we define: Definition 3.3.1 (Generic property). A property P of frameworks is generic in dimension d if whenever some generic realization p of a graph G in Rd has property P, then every generic realization q of G in Rd also has property P. The above discussion about local rigidity and infinitesimal rigidity tells us that local rigidity and infinitesimal rigidity are both generic properties of bar-joint 44 Chapter 3. Rigidity theory frameworks in every dimension. More precisely, given a graph G, if there ex- ists a d-dimensional linearly generic realization (G, p) of G that is locally (resp., infinitesimally) rigid then every d-dimensional generic realization (G, q) of G is also locally (resp., infinitesimally) rigid since the rank of R(G, p) and R(G, q) are equal. Global rigidity is also proved to be a generic property by Gortler, Healy and Thurston [44]. However, universal rigidity is not a generic property (Figure 1.5). When focusing on graphs, we say that a graph G has property P in dimension d if every generic framework of G in dimension d has property P. Note that then locally rigid, infinitesimally rigid and statically rigid mean the same thing for graphs. The study of combinatorial rigidity focuses on finding combinatorial character- izations of the underlying graphs of generic frameworks that have some rigidity property. The first fundamental result in combinatorial rigidity is the famous the- orem of Laman which characterizes the underlying graphs of locally rigid generic frameworks in the plane. We refer to these graphs as generically locally rigid graphs in R2. Theorem 3.3.1 (Laman [75]). A graph G = (V,E) is generically locally rigid in R2 if and only if E contains a subset F that satisfies 1. |F | = 2|V | − 3, and 2. |F ′| ≤ 2|V (F ′)| − 3 for every subset F ′ ⊆ F . In other words, a graph is generically locally rigid in R2 if and only if it contains a (2, 3)-tight spanning subgraph. In many cases, it is convenient to consider a minimally generically locally rigid graph, i.e., a generically locally rigid graph with the property that deleting any edge makes it no more generically locally rigid. These graphs are often referred to as isostatic graphs. The following results are equivalent forms of Laman’s theorem for 2-dimensional isostatic graphs. Theorem 3.3.2 (Lov´asz and Yemini [80], Recski [89]). For a graph G, the following statements are equivalent. 1. G is isostatic in the plane. 2. Duplicating any edge of G results in a graph that is the union of two edge- disjoint spanning trees. 45 3.3. Combinatorial rigidity 3. Adding an edge between any two vertices of G results in a graph that is the union of two edge-disjoint spanning trees. Theorem 3.3.3 (Crapo [26]). A graph G = (V,E) is isostatic in the plane if and only if G can be decomposed into three edge-disjoint trees T1,T2,T3 such that each vertex in V is covered by exactly two of them, and no two subtrees of T1,T2,T3 with more than one vertex span the same set of vertices. A decomposition satisfying the condition in this theorem is call a 3T 2 proper decomposition. From a matroid viewpoint, the generic local rigidity of a graph in the plane can be determined from the rank function of the generic rigidity matroid G2. Theorem 3.3.4 (Lov´asz and Yemini [80]). The rank of G2 is given by rG2 (E) = min{ (2|Vi| − 3) : E(V1),...E(Vm) partition E}. i=1,...,m X One important trend in rigidity theory is to study inductive construction of generically rigid graphs. Complete characterizations of generically locally rigid graphs and generically globally rigid graphs in terms of inductive construction are obtained for bar-joint model in dimension 2. (They are trivial for dimension 1.) This approach is even more fruitful in other models such as body-bar frameworks, body-bar-hinge frameworks, etc., where complete characterizations for generic local rigidity are obtained for all dimensions. Below we summarize results on generic local and global rigidity related to inductive construction. First, we recall the definition of some useful extension operations. Definition 3.3.2 (0-extension, 1-extension). 1.A d-dimensional 0-extension on a graph H is an operation that adds to H a new vertex v and connects v to d different vertices v1, . . . , vd in V (H) (Figure 3.4). 2.A d-dimensional 1-extension on a graph H deletes an existing edge v1v2 in H, adds to H a new vertex v, then connects v to v1, v2 and other d−1 vertices v3, . . . , vd+1 in H (Figure 3.5). A graph G obtained from a graph H by a d-dimensional 0-extension (resp., 1- extension) is called a d-dimensional 0-extension (resp., 1-extension) of H. 0- extension and 1-extension are known to preserve the isostaticity of a graph. 46 Chapter 3. Rigidity theory v • 0-extension • • ... • • • ... • v1 v2 vd v1 v2 vd Figure 3.4: d-dimensional 0-extension. v • 1-extension • • ... • • • ... • v1 v2 vd+1 v1 v2 vd+1 Figure 3.5: d-dimensional 1-extension. Theorem 3.3.5 (see [10, 102]). Suppose that H is a d-dimensional isostatic graph and G is obtained from H by a d-dimensional 0-extension or 1-extension. Then G is d-dimensional isostatic. In dimension 2, these two operations are proved to be sufficient for character- izing isostatic graphs. Theorem 3.3.6 (see, e.g., [10, 47, 58]). A graph G = (V,E) is isostatic in the plane if and only if G can be constructed from K2 by a sequence of 2-dimensional 0-extensions and 1-extensions. (See Figure 3.6.) The problem of characterizing isostatic graphs in dimension 3 seems to be extremely difficult. An example showing that the Laman-type counting condition does not work in dimension 3 is the famous double banana in Figure 3.7. It is easy to verify that, in this graph, for every subset X of V with |X| ≥ 3, i(X) ≤ 3|X|−6 and |E| = 3|V |−6. Yet the graph is not rigid in dimension 3 since there is a relative rotation of the two bananas about the axis through their two ends. A characterization of 3-dimensional isostatic graphs in terms of inductive con- struction may have to consider the following conjecture. Conjecture 3.3.7 (X-replacement [102]). Let G be a 3-dimensional isostatic graph and v1, . . . , v5 distinct vertices of G with v1v2, v3v4 ∈ E(G). Then the operation that deletes v1v2, v3v4 and adds to G a new vertex v and edges vv1 . . . , vv5 results in a 3-dimensional isostatic graph. 47 3.3. Combinatorial rigidity Figure 3.6: A sequence of graphs constructed from K2 by a sequence of 2-dimensional 0- extensions and 1-extensions. Bold edges denote new edges. All these graphs are isostatic. Figure 3.7: The double banana. Although a combinatorial characterization of 3-dimensional generically locally (globally) rigid graphs remains challenging, a positive result is obtained for an important class of graphs: square graphs (equivalently called molecular graphs for their role in modeling molecular structures). The square graph G2 of a graph G is obtained from G by adding an edge between every pair of vertices that have a common neighbor in G (c.f. Figure 1.3). The Molecular Conjecture, proved by Katoh and Tanigawa [72] for body-hinge structures in general dimension, implies the following combinatorial characterization for the 3-dimensional rigidity of square graphs (see [59]). Theorem 3.3.8 (Katoh and Tanigawa [72]). The square graph G2 of a graph G is 48 Chapter 3. Rigidity theory generically locally rigid in dimension 3 if and only if the multigraph 5G, obtained from G by multiplying each edge of G by 5, contains 6 edge-disjoint spanning trees. In dimension one, it is a well known fact that a graph is generically globally rigid if and only if it is 2-connected. The problem of combinatorially characterizing 2-dimensional generically globally rigid graphs was settled by Jackson and Jord´an. A graph is redundantly generically locally rigid in dimension d if deleting any edge results in a generically locally rigid graph in dimension d. Theorem 3.3.9 (Hendrickson [53], Jackson and Jord´an [58]). A graph is generi- cally globally rigid in dimension 2 if and only if it is 3-connected and redundantly generically locally rigid in dimension 2. In fact, the “only if” part of this theorem is shown by Hendrickson [53] for all dimensions. He conjectured the truth of the converse for all dimensions. Jackson and Jord´an confirm it for dimension 2 by proving an inductive construction of graphs that are 3-connected and redundantly locally rigid in the plane: these graphs are constructed from K4 by a sequence of edge-additions and 2-dimensional 1-extensions, which are known to preserve the generic global rigidity. For d = 3, Connelly proved that K5,5 is a counter example. Though infinite families of counter examples for Hendrickson’s conjecture in higher dimensions (d ≥ 5) are obtained 3 [40], it remains an open question whether K5,5 is the only counter example in R . Unlike generic local rigidity and global rigidity, knowledge on combinatorial properties of generically universally rigid graphs is quite modest even in dimension one. Probably, the only known construction to create generically universally rigid graphs is the following. Lemma 3.3.10 (Ratmanski [88]). A graph G on at least d + 2 vertices is d- dimensional generically universally rigid (d-GUR) if G can be obtained from Kd+1 by the following operations: (i) add an edge, (ii) choose two graph G1,G2 built by these operations, choose two sets U1,U2 of each with |U1| = |U2| ≥ d + 1, delete all edges joining vertices of U1 in G1, then glue the two graphs together along the vertices in U1 and U2. In particular, if we add a vertex to a d-GUR graph G and connect it to at least d + 1 vertices of G then we obtain a d-GUR graph. It is an open question whether the converse of Lemma 3.3.10 is true. 49 3.3. Combinatorial rigidity Figure 3.8: A 5-connected graph that is not generically rigid in R2 [80]. An interesting question in combinatorial rigidity theory is that whether high vertex-connectivity implies generic local/global rigidity. For dimension 2, based on Laman’s characterization, Lov´asz and Yemini [80] showed that every 6-vertex- connected graphs minus any three edges are generically locally rigid in R2. They also pointed out that 6 is the minimum value, i.e., there are 5-vertex-connected graphs that are not generically locally rigid in R2 (Figure 3.8). Combining the result of Lov´asz and Yemini with Theorem 3.3.9, it yields that 6-vertex-connected graphs are generically globally rigid in R2. In dimension d ≥ 3, as for characterization of generic local rigidity, the relation between vertex-connectivity and generic local/global rigidity of graphs is unknown. It is conjectured by Lov´asz and Yemini that d(d + 1)-connectivity is sufficient. For molecular graphs, using the known combinatorial characterization proved by Katoh and Tanigawa (Theorem 3.3.8), Jord´an [69] derived that every 7-vertex-connected molecular graphs are generically locally rigid in R3. The same question about the relation between vertex-connectivity and the generical universal rigidity of graphs would be asked. However, in Chapter 8 we prove that no vertex-connectivity can guarantee the generic universal rigidity of graphs in any dimension. In fact, we show that every complete bipartite graph, except K2, is non generically universally rigid in any dimension. 50 Chapter 4 Matroid approach Contents 4.1 Introduction ...... 52 4.2 Characterizing abstract rigidity matroids ...... 53 4.3 1-extendable abstract rigidity matroids ...... 54 4.4 Submodular functions inducing ARMs ...... 58 4.5 A potential application ...... 61 51 4.1. Introduction 4.1 Introduction The concept of abstract rigidity matroid was first introduced by Graver [46] as a generalization of the generic rigidity matroid. Yet application of abstract rigidity matroids in the study of rigidity is still limited, viewing the generic rigidity matroid as an abstract rigidity matroid allows one to concentrate on its combinatorial nature. On the other hand, abstract rigidity matroids are an interesting topic in its own right with many open questions and may have applications in other problems as discussed in Section 4.5. Before going into the definition of abstract rigidity matroids let us keep in mind the two simple but important properties of d-dimensional generic bar-joint frameworks whose detailed proof can be found in [47]. The first one is that, if two frameworks are glued together over at most d − 1 joints, then the composed framework is not rigid (it always allows a relative rotation between the two parts about a d − 2-dimensional affine space containing the glued joints). Moreover, if we add any bar crossing these two parts then the framework becomes “more rigid”. The second one is that, if two rigid frameworks are glued together over at least d joints, then the obtained framework is also rigid. We can formalize these observations as follows. In this chapter we abuse notation K(V ) to denote also the edge set of the complete graph on a vertex set V and Kt to denote the edge set of a complete graph on t vertices. Let Gd(n) be the generic rigidity matroid on the complete graph K = (V,K(V )) and r its rank function. An edge set E ⊆ K(V ) is said to be rigid if cl(E) = K(V (E)), i.e., E spans K(V (E)). In matroid language, the two properties above are restated as follows. (C1) If |V (E) ∩ V (F )| ≤ d − 1, then cl(E ∪ F ) ⊆ K(V (E)) ∪ K(V (F )). (C2) For every pair of rigid subsets E,F of K(V ), if |V (E) ∩ V (F )| ≥ d, then E ∪ F is rigid. Let Ad be a matroid on K(V ) with closure operator clAd (·). A subset E of K(V ) is said to be rigid (in Ad) if clAd (E) = K(V (E)). Then Ad is called a d-dimensional abstract rigidity matroid if it satisfies (C1) and (C2). This chapter presents an approach to the problem of characterizing the generic rigidity matroid from a pure matroid viewpoint. First, in Section 4.2, we describe our result on combinatorial characterization of abstract rigidity matroids. Then, in Section 4.3, we introduce the concept of 1-extendable abstract rigidity matroid – a generalization that captures more properties of the generic rigidity matroid than 52 Chapter 4. Matroid approach abstract rigidity matroids – and show that although in dimension 2 a 1-extendable abstract rigidity matroid coincides with the generic rigidity matroid, in dimension 3 they can be different. Section 4.4 is devoted to the study of intersecting submodular functions that induce abstract rigidity matroids. We provide a necessary condition for these functions. With an additional assumption on the symmetry, we show that this necessary condition is also sufficient. We close the chapter with a discussion on a potential application of our results on abstract rigidity matroids. This work originates from the author’s master’s thesis and partly published in [85] before the enrolment to the PhD program. We include these parts in this thesis to provide a complete view of the early approach. So many details and proofs in these parts will be omitted. 4.2 Characterizing abstract rigidity matroids A vertex star of the complete graph (V,K(V )) is the set of all edges incident to some vertex v ∈ V . In [47], Graver, Servatius and Servatius posed two questions on the characterization of abstract rigidity matroids in dimension 2. Question 1 [47, page 107] Is it true that a matroid M on the edge set of the complete graph (V,K(V )) is a 2-dimensional abstract rigidity matroid if and only if all of the K4’s are circuits and all of the vertex stars minus an edge are cocircuits? Question 2 [47, page 108] Is it true that a matroid M on the edge set of the complete graph (V,K(V )) is a 2-dimensional abstract rigidity matroid if and only if all of the K4’s are circuits and r(K(U)) = 2|U|−3 for all U ⊆ V with |U| ≥ 2? Subsequently, in [48], they gave an affirmative answer to Question 1 together with its generalization in higher dimension. In [85], we show that the condition in Question 1 and the one in Question 2 are both equivalent to the property that M is an abstract rigidity matroid. This gives an affirmative answer to Question 2 and its generalization as well as an alternative proof to the result of Graver, Servatius and Servatius [48, Theorem 0.2]. As a byproduct, we obtain a polynomial algorithm for testing if a matroid given by an independence oracle is a d-dimensional abstract rigidity matroid for any fixed d. For E ⊆ K, v ∈ V \V (E) and u1, . . . , uk ∈ V (E), we call F = E+vu1 +···+vuk a k-valent 0-extension of E. Let Kt denote the edge set of a complete subgraph 53 4.3. 1-extendable abstract rigidity matroids on t vertices of (V,K). The principal ingredient to prove our characterization of abstract rigidity matroids is the following lemma. Lemma 4.2.1 ([85]). A matroid on the edge set of the complete graph (V,K) is a d-dimensional abstract rigidity matroid if and only if it satisfies: 1. rAd (K(V )) = d|V | − d(d + 1)/2; 2. Each k-valent 0-extension of an independent set of Ad is also an independent set of Ad for every k ≤ d. The following theorem answers the two questions above and provides charac- terizations of d-dimensional abstract rigidity matroids for any d ≥ 2. Theorem 4.2.2 ([85]). The following statements are equivalent for a matroid Ad on K(V ). (i) Ad is a d-dimensional abstract rigidity matroid on K. (ii) All Kd+2’s in K(V ) are circuits of Ad and all vertex stars minus (d−1) edges are cocircuits of Ad. (iii) All Kd+2’s in K(V ) are circuits of Ad and rAd (K(U)) = d|U| − d(d + 1)/2 for every U ⊆ V with |U| ≥ d + 1. (iv) All Kd+2’s in K(V ) are circuits of Ad and rAd (K) = d|V | − d(d + 1)/2. Theorem 4.2.2 implies that we can discern whether a matroid Ad given by an independence oracle is a d-dimensional abstract rigidity matroid for a fixed positive integer d in polynomial time by checking condition (iv). Although it is not mentioned in [48], condition (ii) also implies a polynomial time algorithm for testing whether a given matroid is a d-dimensional abstract rigidity matroid. However, to check condition (ii), we would need to verify whether every vertex star minus (d − 1) edges is a cocircuit, or, equivalently, its complement is a hyperplane, which would take O(nd+1) calls to the independence oracle, while an algorithm using condition (iv) would need only O(nd) oracle calls. 4.3 1-extendable abstract rigidity matroids In this section, extensions on a graph are regarded as extensions on its edge set. A matroid M on the edge set K of the complete graph (V,K) is called a d-dimensional 1-extendable abstract rigidity matroid if M is a d-dimensional abstract rigidity 54 Chapter 4. Matroid approach Figure 4.1: The graph H in the proof of Theorem 4.3.1. Each vertex is connected to four nearest vertices and its opposite vertex. |H| = 30 = r(K12) in G3(12). v k+1 vm vk+1 v m v vertex v v v k v’ k splitting v v v1 v2 1 2 Figure 4.2: Vertex splitting operation matroid on K and if F ⊆ K is a d-dimensional 1-extension of an independent set in M then F is independent in M. The generic rigidity matroid on K is an example of a 1-extendable abstract rigidity matroid on K. In dimension 2, the generic rigidity matroid is charac- terized by Laman’s condition and also by Theorem 3.3.6. A corollary of these characterizations is that the 2-dimensional generic rigidity matroid is the unique 2-dimensional 1-extendable abstract rigidity matroid (Graver, Servatius and Ser- vatius [47, Theorem 4.2.3]). The following theorem shows that G3 is not the only 3-dimensional 1-extendable abstract rigidity matroid. Theorem 4.3.1. There exists a 3-dimensional 1-extendable abstract rigidity ma- troid that is not a generic rigidity matroid. We briefly describe the idea of the proof of Theorem 4.3.1. Let us consider the subset H of K = K12 depicted in Figure 4.1. ′ For a subset E ⊂ K with vv1, vv2, . . . , vvm ∈ E and a vertex v ∈ V \ V (E) ′ ′ ′ ′ ′ the edge set F = E + vv + v v1 + v v2 − vvk+1 − · · · − vvm + v vk+1 + ··· v vm with 2 ≤ k ≤ m is said to be obtained from E by a vertex splitting operation (Figure 4.2). In G3, if E is independent, then F is also independent (Whiteley [108]). The graph H can be obtained from K4 by a vertex splitting operation and a 55 4.3. 1-extendable abstract rigidity matroids vertex 1−ext. splitting 1−ext. 1−ext. 1−ext. 0−ext. 0−ext. 0−ext. Figure 4.3: Building H from K4 by a sequence of 0-extensions, 1-extensions and a vertex splitting operation. New vertices are denoted in black. sequence of 1-extensions and 0-extensions as shown in Figure 4.3. Since K4 is in- dependent in the 3-dimensional generic rigidity matroid, E(H) is also independent in G3(12). Furthermore, we can show that the subset H − e + f is a base of K in G3(12) for every e ∈ H and f ∈ K \ H. Figure 4.4 illustrates how a graph H + e − f can be constructed from K4 by 0-extensions and 1-extensions. ′ Now, let B be the set of bases of G3(12), then H ∈ B. Let B = B − H. Then using the fact that E(H) − e + f is a base of K in G3(12) for every e ∈ H and f ∈ K \ E(H) we can show that B′ is a set of bases of a matroid M′ on K. This matroid M′ is obviously a 1-extendable abstract rigidity matroid since the only ′ base that we delete from G3(12) to obtain M is the edge set of a graph with all vertices of degree 5. Remark: Walter Whiteley communicated that, in dimension d ≥ 4, it had been already known that the spline matroid is a 1-extendable abstract rigidity ma- troid which is distinct from the generic rigidity matroid. In dimension 3, however, the spline matroid is conjectured to be isomorphic to the rigidity matroids [110]. Moreover, the example in the proof of Theorem 4.3.1 is also showing that the inde- 56 Chapter 4. Matroid approach 1−ext. 1−ext. 1−ext. 1−ext. 1−ext. 0−ext. 1−ext. 0−ext. Figure 4.4: Building an edge set H − e + f from K4 by a sequence of 0-extensions and 1-extensions. New vertices are denoted by black nodes. 57 4.4. Submodular functions inducing ARMs pendence preservation of vertex splitting operation is not a matroidal consequence of 1-extendability. Using the graph K6,6 minus a perfect matching one can show that the independence preservation of X-replacement operation is not a matroidal consequence 1-extendability and the independence preservation of vertex splitting operation. 4.4 Intersecting submodular functions inducing abstract rigidity matroids. Let us recall that an integer-valued function f : 2S → Z is called an intersecting submodular function if f(X) + f(Y ) ≥ f(X ∪ Y ) + f(X ∩ Y ) for every X,Y ⊆ S with X ∩ Y =6 ∅. In this section when we talk about intersecting submodular functions, we mean non-decreasing intersecting submodular functions. Recall also that an intersecting submodular function f : 2S → Z induces a matroid with the collection of independent sets I(f) = {I ⊆ S | f(J) ≥ |J|, ∀J ⊆ I,J =6 ∅}. We say that a matroid M on S is induced by intersecting submodular function f if the collection of independent sets of M coincides with I(f). Note that different intersecting submodular functions can induce the same matroid. Laman’s condition can be restated as follows. The intersecting submodular function f defined by f(X) = 2|V (X)|−3, for ∅ 6= X ⊆ K, induces G2(n). A natural question is: What are necessary and sufficient conditions for an intersecting submodular function to induce Gd(n)? We can also consider a relaxed version on the conditions for an intersecting submodular function to induce an abstract rigidity matroid. In the following, we will show that all intersecting submodular functions induc- ing an abstract rigidity matroid must have the same value as the rank function of the rigidity matroid on the edge set of a complete subgraph of K(V ). Conversely, this condition together with the “symmetry” will ensure that the induced matroid is an abstract rigidity matroid. Theorem 4.4.1. Let f : 2K → Z be an intersecting submodular function that induces a d-dimensional abstract rigidity matroid M on K. Let Kt ⊆ K be the 58 Chapter 4. Matroid approach edge set of a complete subgraph on t vertices for t ≥ 1. Then f(Kt) = dt−d(d+1)/2 holds for t ≥ d + 2. We say that a function f : 2K → Z is iso-symmetric if f(E) = f(F ) whenever E,F ⊆ K induce two isomorphic subgraphs. K Theorem 4.4.2. Suppose that f : 2 → Z satisfies f(Kt) = dt − d(d + 1)/2 for t ≥ d + 2, f(Kt) ≥ |Kt| for t ≤ d + 1, for the edge set Kt of any complete subgraph on t vertices of K. Suppose further that f is iso-symmetric. Then f induces a d-dimensional abstract rigidity matroid on K. For the sake of simplicity we will demonstrate the two theorems above for the 2-dimensional case. The same arguments work for higher dimensions. Proof of Theorem 4.4.1. We prove by induction on t. Let K4 ⊆ K be the edge set of a complete subgraph on 4 vertices. Since K4 is a circuit of the abstract rigidity matroid M (by Theorem 4.2.2) and f induces M we have f(K4) < |K4| and f(K4 − e) ≥ |K4 − e| for all e ∈ K4. Therefore, f(K4) = |K4|−1 = 2×4−3 holds. Suppose that f(Kt) = 2t−3 holds for some t ≥ 4. Claim 4.4.3. Let v1, v2 ∈ V (Kt), v ∈ V (K)\V (Kt). Then f(Kt + vv1 + vv2) = f(Kt) + 2. Proof. We have f(Kt + vv1 + vv2) ≥ rM(Kt + vv1 + vv2) = rM (Kt) + 2 (by Lemma 4.2.1) = f(Kt) + 2 (by induction hypothesis). Suppose on contrary that f(Kt + vv1 + vv2) ≥ f(Kt) + 3. Let v3 be a vertex in V (Kt) − {v1, v2} and Bt a base of Kt. We prove that Bt + vv1 + vv2 + vv3 is independent in M, which has greater rank than Kt+1, a contradiction. We just need to prove that f(X) ≥ |X| for every subset X of Bt + vv1 + vv2 + vv3. If {vv1, vv2, vv3} * X or X = {vv1, vv2, vv3} then evidently X is independent in M by Lemma 4.2.1, thus f(X) ≥ |X| holds. Now, if X = Xt + vv1 + vv2 + vv3 with ∅ 6= Xt ⊆ Bt, then f(X) + f(Bt) ≥ f(X ∪ Bt) + f(X ∩ Bt) = f(Bt + vv1 + vv2 + vv3) + f(Xt) ≥ f(Bt) + 3 + |Xt|. 59 4.4. Submodular functions inducing ARMs It implies that f(X) ≥ |X| holds. The following claim completes the proof of Theorem 4.4.1 for d = 2. Claim 4.4.4. Let k be an integer such that 2 ≤ k ≤ t. Suppose that v1, . . . , vk ∈ V (Kt) and v∈ / V (Kt). Then f(Kt + vv1 + ··· + vvk) = f(Kt) + 2 holds. Proof. We prove by induction on k. When k = 2 the statement holds by Claim 4.4.3. For k = 3, if f(Kt + vv1 + vv2 + vv3) ≥ f(Kt) + 3 then using the same argument as that in Claim 4.4.3 we deduce a contradiction. Thus the statement holds for k = 3. Now suppose that the statement holds for some k ≥ 3. Let vk+1 be a vertex in V (Kt) − {v1, . . . vk}, then using the intersecting submodularity of f we have f(Kt + vv1 + vv2 + ··· + vvk + vvk+1) + f(Kt + vv2 + ··· + vvk) ≤ f(Kt + vv1 + ··· + vvk) + f(Kt + vv2 + ··· + vvk+1), which implies that f(Kt + vv1 + vv2 + ··· + vvk + vvk+1) ≤ f(Kt) + 2 and thus f(Kt + vv1 + vv2 + ··· + vvk + vvk+1) = f(Kt) + 2. Proof of Theorem 4.4.2. Let M(f) be the matroid induced by f. Let Kt be the edge set of an arbitrary complete subgraph on t vertices of K. Claim 4.4.5. If v∈ / V (Kt) and v1 ∈ V (Kt) then f(Kt + vv1) ≥ f(Kt) + 1 holds. Proof. Assume on the contrary that f(Kt + vv1) = f(Kt). Then, by the iso- symmetry of f, f(Kt + vvi) = f(Kt) holds for every vi ∈ V (Kt). Using the intersecting submodularity of f we can easily derive that f(Kt) = f(Kt + vv1 + ··· + vvt), where v1, . . . , vt are all the vertices of Kt. Then the edge set Kt+1 = Kt +vv1 +···+vvt of the complete subgraph on t+1 vertices satisfies 2(t+1)−3 = f(Kt+1) = f(Kt) = 2t − 3, which is a contradiction. Claim 4.4.6. If v∈ / V (Kt) and v1, v2 ∈ V (Kt) then f(Kt +vv1 +vv2) ≥ f(Kt)+2. Proof. From Claim 4.4.5 we have f(Kt + vv1 + vv2) ≥ f(Kt) + 1. Assume on the contrary that f(Kt +vv1 +vv2) = f(Kt)+1. Using induction on k, the intersecting submodularity and the iso-symmetry of f we can deduce that f(Kt + vv1 + ··· + vvk) = f(Kt) + 1 holds for every k and v1, . . . , vk ∈ V (Kt). In particular, the edge set Kt+1 = Kt + vv1 + ··· + vvt of the complete subgraph on t + 1 vertices satisfies 2(t + 1) − 3 = f(Kt+1) = f(Kt) + 1 = 2t − 2, a contradiction. 60 Chapter 4. Matroid approach Claim 4.4.7. A 1-valent 0-extension of an independent set of M(f) is again an independent set of M(f). Proof. Let Bt be a base of Kt in matroid M(f), v∈ / V (Kt) and v1 ∈ V (Kt). Then f(Bt + vv1) = f(Kt + vv1) ≥ f(Kt) + 1 = f(Bt) + 1, by Claim 4.4.5. Let X be an arbitrary non-emty subset of Bt. By the intersecting submodularity f(X + vv1) + f(Bt) ≥ f((X + vv1) ∪ Bt) + f((X + vv1) ∩ Bt) = f(Bt + vv1) + f(X) ≥ f(Bt) + 1 + |X|. Thus, f(X + vv1) ≥ |X + vv1| holds. It follows that Bt + vv1 is independent in the matroid M(f), which implies the statement of Claim 4.4.7 Claim 4.4.8. A 2-valent 0-extension of an independent set in M(f) is again an independent set in M(f). Proof. Let Bt be a base of Kt in the matroid M(f), v∈ / V (Kt) and v1, v2 ∈ V (Kt). Assume on the contrary that there exists a subset X of Bt + vv1 + vv2 such that X is a circuit in M(f). Then, by Claim 4.4.7, X = Xt + vv1 + vv2 with Xt ⊆ Bt. Then, by the intersecting submodularity of f, f(X) + f(Bt) ≥ f(X ∪ Bt) + f(X ∩ Bt) = f(Bt + vv1 + vv2) + f(Xt) = f(Kt + vv1 + vv2) + f(Xt) ≥ f(Bt) + 2 + |Xt| (by Claim 4.4.6). Therefore, f(X) ≥ |X| holds, a contradiction. Claim 4.4.7, Claim 4.4.8 and Theorem 4.2.2 imply that M(f) is a 2-dimensional abstract rigidity matroid on K, which completes the proof of Theorem 3.7 for d = 2. 4.5 A potential application We end this chapter by discussing a potential application of our results on abstract rigidity matroids. Thomassen [103] conjectured that there exists a function f(k) such that every f(k)-connected graph has a k-connected orientation. Jord´an [68] confirmed this 61 4.5. A potential application conjecture for k = 2 by showing that f(2) ≤ 18. This upper bound is improved by Cheriyan, Durand de Gevigney and Szigeti [18] to f(2) ≤ 14, using a similar idea. For k ≥ 3 no upper bound has been obtained. The main idea of Jord´an [68] is that a 18-connected graph contains 3 edge-disjoint 2-connect spanning subgraphs. Using these 2-connected subgraphs he deduces a 2-connected orientation of the original graph. Thus comes the question if we can pack many k-connected spanning subgraphs in a highly connected graph. The 2-connected graphs used by Jord´an are in fact 2-dimensional generically rigid graphs and the packing is obtained by using the rank formula given by the intersecting submodular function f(F ) = 2|V (F )| − 3 which induces G2. Therefore, we are interested in the question of finding a matroid such that the “rigid” graphs with respect to this matroid are k-connected, and that it possesses a “simple” inducing intersecting submodular function. k-dimensional abstract rigid- ity matroids may be good candidates since rigid graphs with respect to these ma- troids are k-connected and the sufficient condition for an intersecting submodular function to induce an abstract rigidity matroid is simple as shown in Section 4.4. 62 Chapter 5 Inductive constructions and decompositions of graphs Contents 5.1 Introduction ...... 65 5.2 Graded sparse graphs ...... 68 5.2.1 Introduction ...... 68 5.2.2 A reduction theorem for graded sparse graphs ...... 70 5.2.3 Inductive construction of graded sparse graphs . . . . . 75 5.2.4 Decomposition of graded sparse graphs ...... 75 5.2.5 Graded sparse matroids ...... 81 5.3 (b, l)-sparse graphs ...... 86 5.3.1 Inductive construction ...... 86 5.3.2 (b, l)-pebble games ...... 93 5.4 Packing of matroid-based arborescences ...... 96 5.4.1 Introduction ...... 96 5.4.2 Proof of the main theorem ...... 102 5.4.3 Polyhedral description ...... 105 5.4.4 Algorithmic aspects ...... 106 63 5.4.5 Further remarks ...... 107 64 Chapter 5. Inductive constructions and decompositions 5.1 Introduction This chapter presents combinatorial optimization results developed to solve the problem of characterizing the underlying graphs of infinitesimally rigid generic frameworks of many types. The starting point is the following fundamental result of Nash-Williams. Theorem 5.1.1 (Nash-Williams [83]). A graph G = (V,E) can be decomposed into m edge-disjoint spanning trees if and only if i(X) ≤ m|X| − m for every non empty subset X of V and |E| = m|V | − m. Recall that a graph G verifying the condition in the theorem above is called an (m, m)-tight graph. An inductive construction for (m, m)-tight graphs is also implicit in [83]. In fact, a typical proof that a graph has an infinitesimally rigid realization if and only if it has a spanning tight sparse subgraph, proceeds along the following lines: (a) Construct a rigidity matrix whose rank determines the infinitesimal rigidity of a framework; (b) Deduce that the existence of a tight sparse subgraph is a necessary condition for infinitesimal rigidity; (c) Use either an inductive construction, or a decomposition, of a tight sparse subgraph to demonstrate sufficiency by constructing a realization whose rigid- ity matrix attains the maximum rank. An inductive construction for (2, 3)-sparse graphs suggested by Henneberg is used to prove Laman’s fundamental result (see [10, 47, 58]. Similarly, Tay [99] used the inductive construction for (m, m)-tight graphs due to Nash-Williams [82], to d+1 show that when m = 2 , these graphs are exactly the underlying graphs of min- imally infinitesimally rigid d-dimensional generic body-bar frameworks. Inductive techniques were also employed successfully by Katoh and Tanigawa [72] to settle the long-standing Molecular Conjecture (on rigidity of panel-hinge frameworks). The alternative approach in (c) is to use a decomposition of tight sparse graphs to give a direct construction of an infinitesimally rigid realization. Examples of this approach are Tay [98], Whiteley [107], and Jackson and Jord´an [60]. The above mentioned result of Nash-Williams [82] was not motivated by rigidity theory. However, the applications mentioned above have stimulated interest in generalizing the results of Nash-Williams and Tutte. 65 5.1. Introduction The first way to generalize these results is to consider (m, l)-tight graphs. Fekete and Szeg˝o[33] provide an inductive construction of (m, l)-tight graphs for 0 ≤ l ≤ m. Whiteley [109] deduces a decomposition of (m, l)-tigh graph for 0 ≤ l ≤ m from matroid decomposition. Haas [51] characterizes (m, l)-tight sparse graph for 0 ≤ m ≤ l < 2m − 1 in terms of an “lT m-tree decomposition”, a concept rooted in Crapo’s work on decomposition of Laman’s graphs [26]. These decompositions are re-obtained by Streinu and Theran [96] using pebble games. The study of frameworks with different kinds of constraints led Lee, Streinu and Theran [79] to consider a more generalized class of graphs in which different types of edges satisfy different sparsity conditions. They call these graphs graded sparse graphs. As an example they consider the bar-joint-slider model in which bars are represented by non-loop edges and sliders are represented by loops. They show that a graph can be realized as a rigid slider-bar-joint 2-dimensional framework if and only if it has a subgraph H which is (2, 0)-tight sparse and the subgraph of H induced by the non-loop edges is (2, 3)-sparse. In section 5.2, we provide an inductive construction as well as a decomposition for graded tight sparse graphs. The decomposition will be used in Chapter 6 to characterize different types of body- length-direction frameworks as well as to re-obtain a result of Katoh and Tanigawa on the characterization of underlying graphs of generic body-bar frameworks with bar-boundary. We also propose a second way to generalize the result of Nash-Williams by considering (b, l)-tight sparse graphs. These tight sparse graphs arise in the rigidity context when one considers frameworks that contain bodies of different dimensions. An inductive construction for a special instant of b, l was derived by Tay [98] to characterize body-rod-bar frameworks. In section 5.3, we give an inductive construction of (b, l)-tight graphs for l not greater than the minimum value of b. We also characterize (b, l)-sparse graphs as resulting graphs in (b, l)-pebble games, a generalization of (k, l)-pebble games by Lee, Streinu [78]. A third way of generalization is proposed by Katoh and Tanigawa in [73]. They regard each tree as a tree with a root r and replace the condition that each vertex is covered by all the m trees 1 with a matroidal condition on the root set of trees covering each vertex. Motivated by the study of bar-joint-slider frameworks, their result is applied successfully to characterize several other types of frameworks with boundaries (cf. Section 6.2). In Section 5.4, we derive a directed counterpart to 1This condition is equivalent to the condition that the m trees in the decomposition/packing are spanning. 66 Chapter 5. Inductive constructions and decompositions the result of Katoh and Tanigawa. Furthermore, we show that our directed result implies their undirected result. As a consequence, we obtain a shorter proof for Katoh and Tanigawa’s result. 67 5.2. Graded sparse graphs 5.2 Graded sparse graphs 5.2.1 Introduction Let G = (V,E) be a graph. An r-grading of E is a strictly decreasing sequence of sets (E1,...,Er) with E = E1 ⊃ E2 · · · ⊃ Er. A graph G with an r-grading (E1,...,Er) is called an r-graded graph, or simply a graded graph when the value of r is clear. The grade of an edge e ∈ E, denoted by σe, is the maximum number t such that e ∈ Et. Let m be a positive integer and d = (d1, . . . , dr) be an r-tuple of integers with 0 ≤ d1 ≤ · · · ≤ dr ≤ 2m − 1. The r-graded graph G is said to be (m, d)-graded sparse if the subgraph Gt = (V,Et) is (m, dt)-sparse for every t = 1, . . . , r, and (m, d)-graded tight if in addition it has |E| = m|V | − d1. For a subset X of V , we denote by it(X) the number of edges of Et induced by X. The graded sparsity condition can then be written as it(X) ≤ m|X| − dt, for all non-empty X ⊆ V and all t = 1, . . . , r. We will need some more definitions and notations to state our main results on graded sparse graphs. A splitting off at a vertex v of G is the operation which deletes two edges f = vx and g = vy incident to v and adds a new edge e = xy of grade σe = min{σf , σg}. When f is a loop at v the splitting off just deletes f and changes the grade of g to min{σf , σg}. We will refer to such a splitting off as a loop replacement. A splitting off for which neither f nor g is a loop will be referred to as a proper splitting off. A(k, ℓ)-reduction of an (m, d)-graded tight graph G is defined as follows. We first choose a vertex v which is incident with m+k edges in total, including exactly ℓ loops, for some 0 ≤ k ≤ m − ℓ. If k = 0 then we simply delete v from G. If k ≥ 1 then we perform a sequence of ℓ loop replacements followed by k proper splitting offs at v, then delete v and the remaining m − k − ℓ edges incident to it. A (k, ℓ)-reduction at v is admissible if the resulting graph is also (m, d)-graded tight. The inverse operation to a (k, ℓ)-reduction is called a (k, ℓ)-extension. A (k, ℓ)- extension of H is an r-graded graph obtained from H by the following steps. (a) Delete k edges ei = xiyi, 1 ≤ i ≤ k, from H. (b) Add a new vertex v and m − k − ℓ new edges of arbitrary grades from v to the vertices of H. 68 Chapter 5. Inductive constructions and decompositions (c) For each i, 1 ≤ i ≤ k, add new edges fi = vxi, gi = vyi and ℓi new loops hi,j at v in such a way that: (i) the minimum grade of fi, gi, hi,1, . . . , hi,ℓi is equal to σei ; (ii) ℓ1 + ℓ2 + ... + ℓk = ℓ; (iii) the total number of loops of grade at least t added at v is less than or equal to m − dt for all 1 ≤ t ≤ r. In the special case when k = 0 we obtain a (0, ℓ)-extension by simply adding a new vertex v, ℓ loops incident with v and m − ℓ edges from v to H in such a way that condition (c)(iii) above is satisfied. Contribution: Our first result is that admissible reductions always exist, when 0 ≤ d1 ≤ · · · ≤ dr ≤ m. Theorem 5.2.1. Suppose that m is a positive integer and d = (d1, . . . , dr) is a non-decreasing sequence of integers 0 ≤ d1 ≤ · · · ≤ dr ≤ m. Let G be an (m, d)- graded tight graph on at least two vertices. Then G has a vertex v which is incident with m + k edges, including exactly ℓ loops, for some 0 ≤ k ≤ m − ℓ. Furthermore, for any such vertex v, G has an admissible (k, ℓ)-reduction at v. We will show that a (k, ℓ)-extension preserves the property of being (m, d)- graded tight. Combined with Theorem 5.2.1, this will imply the following inductive construction for (m, d)-graded tight graphs. Theorem 5.2.2. Suppose that m is a positive integer and d = (d1, . . . , dr) a non- dereasing sequence of integers 0 ≤ d1 ≤ · · · ≤ dr ≤ m. Let G be an r-graded graph. Then G is (m, d)-graded tight if and only if G can be obtained from an (m, d)- graded tight graph on one vertex by a sequence of (k, ℓ)-extensions. Moreover, if d1 > 0 then we need only (k, ℓ)-extensions with k + ℓ < m. A pseudoforest is a graph in which each connected component contains at most one cycle. A pseudoforest is tight if each of its components contains exactly one cy- cle. (Equivalently G is a pseudoforest if it is (1, 0)-sparse and is a tight pseudoforest if it is (1, 0)-tight.) An (m, d)-graded pseudoforest decomposition of an r-graded graph G is a partition of E(G) into m edge-disjoint pseudoforests F1,F2,...,Fm such that for each t, 1 ≤ t ≤ r, and all i, 1 ≤ i ≤ dt, the restriction of Fi on Et is a forest. Our third main result characterizes (m, d)-graded tightness in terms of (m, d)-graded pseudoforest decompositions. 69 5.2. Graded sparse graphs Theorem 5.2.3. Let 0 ≤ d1 ≤ · · · ≤ dr ≤ m be integers and G an r-graded graph. Then G is (m, d)-graded tight if and only if G has an (m, d)-graded pseud- oforest decomposition consisting of d1 spanning trees and m − d1 spanning tight pseudoforests. Theorems 5.2.1, 5.2.2 and 5.2.3 will be proved in Sections 5.2.2, 5.2.3 and 5.2.4, respectively. Furthermore, in Section 5.2.5 we study the matroids whose independent sets are the edge sets of (m, d)-graded sparse graphs. We give a short direct proof for their matroid decomposition and thus obtain an alternative proof for the pseudo- forest decomposition in Theorem 5.2.3. Results in this section are from a joint work with Bill Jackson [66]. 5.2.2 A reduction theorem for graded sparse graphs We suppose throughout this section that m is a positive integer and d = (d1, . . . , dr) with 0 ≤ d1 ≤ · · · ≤ dr ≤ m. (5.2.1) Our aim is to prove Theorem 5.2.1. We first show that every (m, d)-graded tight graph G has a vertex which is incident with the required number of edges. Lemma 5.2.4. Let G = (V,E) be an (m, d)-graded tight graph. Then G has a vertex v which is incident with m + k edges for some 0 ≤ k ≤ m − i(v). Proof. Let v be a vertex of minimum degree in G. Since G is tight, we have d(x) = 2|E| = 2m|V | − 2d1 ≤ 2m|V |. x∈V X Hence d(v) ≤ 2m and so d(v) − i(v) ≤ m + m − i(v). On the other hand m|V | − d1 = |E| = i(V − v) + d(v) − i(v) ≤ m(|V | − 1) − d1 + d(v) − i(v) so d(v) − i(v) ≥ m. We next show that if v is a vertex of G with d(v) − i(v) = m + k for some 1 ≤ k ≤ m − i(v), then we can construct an (m, d)-graded sparse graph from G by performing a sequence of splitting offs at v (consisting of i(v) loop replacements followed by k proper splitting offs). We construct these splitting offs one at a time 70 Chapter 5. Inductive constructions and decompositions in such a way that we preserve (m, d)-graded sparseness as well as an additional sparsity condition at v. We assume for the remainder of this section that H = (V,E) is an (m, d)-graded sparse graph with grading (E1,...,Er) (5.2.2) and v is a vertex of H with d(v) = m + k + c − 2s, i(v) = max{0, c − s}, c ≥ 0, k ≥ 1 and 0 ≤ s < k + c ≤ m. (5.2.3) We imagine that c is the number of loops incident to v in G and that H has been obtained from G by performing s splitting offs at v. Noting that the number of edges incident to v is d(v) − i(v), condition (5.2.3) implies that The number of edges incident to v is strictly greater than m − s. (5.2.4) In fact, if i(v) = 0 then the number of edges incident to v is d(v) = m+k +c−2s = m − s + (k + c − s) > m − s by (5.2.3). If i(v) > 0 then, by (5.2.3), i(v) = c − s, so the number of edges incident to v is d(v) − i(v) = m + k + c − 2s − (c − s) = m + k − s > m − s since k ≥ 1. For X,Y ⊆ V , we denote the intersection of Et and the set of edges from X to Y by Et(X,Y ). (This set contains in particular all edges of Et induced by X ∩ Y .) We also use E(X,Y ) for E1(X,Y ), it(X,Y ) for |Et(X,Y )| and i(X,Y ) for |E(X,Y )|. We say that H is (v, s)-good if it(X) ≤ m|X| − dt − s for all X ⊆ V that properly contain v, and all 1 ≤ t ≤ r. We will suppose henceforth that H is (v, s)-good. (5.2.5) Our aim is to find a splitting off at v in H such that the new graph is both (m, d)-graded sparse and (v, s+1)-good. We call such a splitting off feasible. To do this we need to consider the circumstances in which a splitting off is not feasible. A nonempty set X ⊆ V − v is said to be t-critical, or simply critical, if it satisfies it(X) = m|X| − dt for some 1 ≤ t ≤ r. A set X ⊆ V is t-crucial, or simply crucial, if it properly contains v and satisfies it(X) = m|X| − dt − s. The grade σ(X) of a crucial set X is the maximum value of t for which X is t-crucial. The following result characterizes when the splitting off operation is feasible. Lemma 5.2.5. Let f = vx and g = vy be two edges incident to v with σf ≤ σg. Then splitting off f and g at v is feasible if and only if (a) no t-critical set contains both x and y for all 1 ≤ t ≤ σf , and 71 5.2. Graded sparse graphs (b) all crucial sets X have σ(X) ≤ σg and X ∩ {x, y}= 6 ∅. In addition, if σ(X) > σf , then y ∈ X. Proof. This follows from the definitions of a feasible splitting off, and critical and crucial sets. It follows that we will need to analyze the structure of the families of critical and crucial sets in order to show that there exists a feasible splitting off. Our basic tool is the following result. Lemma 5.2.6. Suppose that X,Y ⊆ V and that t, t′ are integers with 1 ≤ t ≤ t′ ≤ r. Then it(X) + it′ (Y ) + it(X \ Y,Y \ X) ≤ it(X ∪ Y ) + it′ (X ∩ Y ). Proof. This follows easily by counting the contribution of each edge of H to both sides of the inequality. Lemma 5.2.7. Suppose X is a t-critical set, Y is a t′-critical set, t ≤ t′ and ′ X ∩ Y =6 ∅. Then X ∪ Y is t-critical, X ∩ Y is t -critical and it(X \ Y,Y \ X) = 0. Proof. Since H is (m, d) graded sparse, we have it(X ∪ Y ) + it′ (X ∩ Y ) ≤ m|X ∪ Y | − dt + m|X ∩ Y | − dt′ = m|X| − dt + m|Y | − dt′ = it(X) + it′ (Y ) ≤ it(X ∪ Y ) + it′ (X ∩ Y ) − it(X \ Y,Y \ X). Hence equality must occur everywhere and so it(X ∪ Y ) = m|X ∪ Y | − dt, it′ (X ∪ Y ) = m|X ∪ Y | − dt′ and it(X \ Y,Y \ X) = 0. An immediate consequence of Lemma 5.2.7 is the following. Corollary 5.2.8. Suppose x ∈ V −v is contained in at least one critical set. Then the union of all critical sets which contain x is a t-critical set where t is the smallest grade of all the critical sets which contain x. The proof of the following lemma is similar to that of Lemma 5.2.7. Lemma 5.2.9. Suppose that X is a t-crucial set, Y is a t′-crucial set, t ≤ t′ and |X ∩ Y | ≥ 2. Then X ∪ Y is t-crucial and X ∩ Y is t′-crucial. 72 Chapter 5. Inductive constructions and decompositions Lemma 5.2.10. If X is a t-critical set then it(v, X + v) ≤ m − s. Proof. Since H is (v, s)-good, we have m(|X| + 1) − dt − s ≥ it(X + v) = it(X) + it(v, X + v) = m|X| − dt + it(v, X + v). Therefore, it(v, X − v) ≤ m − s. Lemma 5.2.11. If X is a t-crucial set then it(v, X) ≥ m−s and it(v, X −v) ≥ 1. Proof. We have it(X−v)+it(v, X) = it(X) = m|X|−dt−s = m|X−v|−dt+m−s ≥ it(X−v)+m−s since H is (m, d)-graded sparse. Hence it(v, X) ≥ m−s. Suppose it(v, X −v) = 0. Then m − s ≤ it(v, X) = it(v) ≤ i(v) = max{0, c − s}. Since m − s > 0 by (5.2.3), we have m − s ≤ c − s and m ≤ c. This contradicts the facts that m ≥ k + c and k ≥ 1 by (5.2.3). A crucial set X is a minimal crucial set if it is not properly contained in any other crucial set. Lemma 5.2.12. Suppose X and Y are distinct minimal crucial sets of grades t ′ ′ and t , respectively, where t ≤ t . Then X∩Y = {v}, E(v, V ) = Et(v, X)∪Et′ (v, Y ) and E(v) = Et′ (v). Proof. The fact that X ∩ Y = {v} follows from Lemma 5.2.9 and minimality. By Lemma 5.2.11 and (5.2.3), m+k+c−2s = d(v) ≥ it(v, X)+it′ (v, Y ) ≥ 2(m−s) = m+m−2s ≥ m+k+c−2s. Hence, equality must hold everywhere. In particular d(v) = it(v, X) + it′ (v, Y ), which implies that E(v, V ) = Et(v, X) ∪ Et′ (v, Y ) and E(v) = Et′ (v). Lemma 5.2.13. There are at most two distinct minimal crucial sets in H, and the grade of a minimal crucial set is the maximum grade among all the crucial sets that contain it. Proof. Suppose that X1,X2 and X3 are three distinct minimal crucial sets. Then Xi ∩ Xj = {v} and Xi ∪ Xj contains all neighbours of v for all 1 ≤ i < j ≤ 3 by Lemma 5.2.12. This implies that all edges incident to v are loops, and contradicts the fact that i(v, Xi − v) ≥ 1 by Lemma 5.2.11. The second part of the lemma follows immediately from Lemma 5.2.9. 73 5.2. Graded sparse graphs Lemma 5.2.14. There exists a feasible splitting off at v, and this can be taken to be a loop replacement when i(v) ≥ 1. Proof. We first consider the case when i(v) ≥ 1. Let f be a loop at v. Suppose that H has two minimal crucial sets X and Y , of grades t and t′, respectively, ′ where t ≤ t . Then f ∈ Et′ (v) by Lemma 5.2.12 and all crucial sets have grade at ′ most t by Lemma 5.2.13. We may choose an edge g ∈ Et′ (v, Y − v) by Lemma 5.2.11. Lemma 5.2.5 now tells us that splitting off f and g at v will be feasible. We next suppose that H has a unique minimal crucial set X, of grade t say. Then all crucial sets have grade at most t by Lemma 5.2.13 and it(v, X − v) ≥ 1 by Lemma 5.2.11. Lemma 5.2.5 now tells us that splitting off f with any edge g ∈ Et(v, X −v) will be feasible. When H has no crucial sets, Lemma 5.2.5 implies that splitting off f with any other edge at v will be feasible. Hence we may suppose that i(v) = 0. Let us consider the case that H has two minimal crucial sets X and Y of grades ′ t and t , respectively.We may choose edges f = vx ∈ Et(v, X − v) and g = vy ∈ Et′ (v, Y − v) by Lemma 5.2.11. If there is no critical set containing both x and y, splitting off f, g at v is feasible by Lemma 5.2.5. Suppose that there is a critical set containing both x and y. Let U be the maximal one and let t′′ be the minimal integer such that U is t′′-critical. Note that by Corollary 5.2.8, every critical set which contains x is a subset of U and t′′ is the smallest grade of all the critical sets containing x. By Lemma 5.2.10, it′′ (v, U + v) ≤ m − s. On the other hand, the number of edges incident to v is strictly greater than m − s by (5.2.4). Therefore, ′′ we can find an edge h ∈ E(v, V − v) with either σh < t or h ∈ E(v, V − U − v). This edge must belong to Et(v, X − v) or Et′ (v, Y − v) by Lemma 5.2.12. Without ′′ loss of generality, suppose that h ∈ Et′ (v, Y − v). If σh < t then splitting f, h at v is feasible by Lemma 5.2.5. If h = vz ∈ E(v, V − U − v) ∩ Et′ (v, Y − v), we claim that there is no critical set containing both x and z. In fact, if there is a critical set U ′ containing both x and z, by Lemma 5.2.7, U ∪ U ′ is a critical set properly containing x and y, which contradicts the maximality of U. Therefore, splitting off f, h at v is also feasible in this case by Lemma 5.2.5. We now consider the case that there exists at most one minimal crucial set. If there exists a crucial set, let X be the unique minimal crucial set and t its grade. Otherwise let X = V and t = 1. Then by Lemma 5.2.13, there is no crucial set of grade strictly greater than t. Let f = vx be an edge in Et(v, X − v). If there is a critical set containing x, let U be the maximal one and suppose that 74 Chapter 5. Inductive constructions and decompositions t′ is the smallest integer such that U is t′-critical. Note that then there is no t′′ ′′ ′ critical set containing x with t < t . By Lemma 5.2.10, it′ (v, U + v) ≤ m − s. On the other hand, the number of edges incident to v is strictly greater than m − s ′ by (5.2.4). Therefore, we can find an edge h ∈ E(v, V − v) with either σh < t or h ∈ E(v, V − U − v). In both cases, splitting off f, h at v is feasible by the maximality of U and by Lemma 5.2.5. 5.2.3 Inductive construction of graded sparse graphs We assume that (5.2.1) continues to hold throughout this section. Suppose that H is an (m, d)-graded tight graph and k, ℓ are non-negative integers with k + ℓ ≤ m. Lemma 5.2.15. Suppose H is an (m, d)-graded tight graph and G is a (k, ℓ)- extension of H. Then G is (m, d)-graded tight. Proof. We have |E(G)| = |E(H)| + m = m|V (H)| − d1 + m = m|V (G)| − d1. Hence it will suffice to show that G is (m, d)-graded sparse. Choose X ⊆ V (G) G H and t ∈ {1, 2, . . . , r}. If v 6∈ X then it (X) ≤ it (X) ≤ m|X| − dt since H is G (m, d)-graded sparse. If X = {v} then it (X) ≤ m − dt by condition (c)(iii) above. Hence we may suppose that v is properly contained in X. We adopt the notation used in the earlier definition of a (k, ℓ)-extension. Let bt be the number of edges H of {e1, e2, . . . , ek} which are contained in Et (X − v). These edges are deleted in G H Step (a) and hence it (X − v) = it (X − v) − bt ≤ m|X − v| − dt − bt. We add at most m − k − ℓ edges of Et at v in Step (b) and at most 2k + ℓ − (k − bt) edges of Et at v in Step (c). Hence G it (X) ≤ m|X − v| − dt − bt + m − k − ℓ + 2k + ℓ − (k − bt) = m|X| − dt. Since this holds for all t, G is (m, d)-graded sparse. Proof of Theorem 5.2.2 Sufficiency follows from Lemma 5.2.15. Necessity follows by induction on |V (G)| and Theorem 5.2.1. 2 5.2.4 Decomposition of graded sparse graphs In this section we give a proof of the decomposition of (m, d)-graded sparse graphs in Theorem 5.2.3 using the inductive construction. In Section 5.2.5 we give another 75 5.2. Graded sparse graphs proof of this result from matroid decomposition (Theorem 5.2.20). Since this graph decomposition is in fact equivalent to the matroid decomposition, the proof in this section serves also as an alternative proof for the matroid decomposition in Theorem 5.2.17. We will use Theorem 5.2.2 to construct this graph decomposition recursively. Let e = xy be an edge of an r-graded graph H and v be a vertex (not necessary in H). We call a v-subdivision of e the operation that deletes e from H and adds to H two edges f = vx, g = vy with grades σf , σg satisfying σe = min{σf , σg}. The edges f, g are called subdividing edges of e. We call a v-subdivision proper if both f and g are non loops, otherwise we call it improper. Then the (k, ℓ)-extension can be seen as a sequence of k proper v-divisions and ℓ improper v-divisions followed by the addition of m − k − ℓ non loop edges from v to H. Proof of Theorem 5.2.3 We prove by induction on the number of vertices. The decomposition is trivial when |V | = 1. Suppose that G is obtained from H by a sequence of k proper v-subdivisions and ℓ improper v-subdivisions followed by the addition of m − k − ℓ non loop edges from v to H. Our construction of the decomposition of G starts with a pseudoforest decomposition F of H, a set D = {e1, . . . , eh} of edges to be subdivided and a set A of edges which can be added to G. In fact, A = {f1, g1, . . . , fh, gh, a1, . . . , am−h} where fi, gi are subdividing edges of ei, for i = 1, . . . , h. The sets F, A, D will be updated at each v-subdivision and edge addition. At a division of an edge e to f and g, we replace the pseudoforest F ∈ F that contains e with a pseudoforest F ′ by deleting e from F and adding to F some edges from A. We then remove the used edges from A and remove e from D. In fact, if the pseudoforest F has not already covered v then two edges from A are added to F , otherwise, only one is added. At an edge addition of an edge a, we simply choose an appropriate pseudoforest F from F and add a to F . Throughout the remaining of this section we suppose that we have a collection of m graded pseudoforests F = {F1,...,Fm}, a set D of edges to be subdivided and a set A containing the subdividing edges of edges in D as well as other edges to be added. Our construction of the decomposition will decrease |D| + |A|. We will need some more notation. Let Ft denote {Fdt+1,...,Fm} for 1 ≤ t ≤ r. t Set F0 = F and Fr+1 = ∅. For 1 ≤ t ≤ r, we will use A to denote the subset of edges in A of grade at least t and for a pseudoforest F , we also use F t to denote the subgraph of F induced by the edges of grade at least t. When t = r + 1, At 76 Chapter 5. Inductive constructions and decompositions t and F are just empty sets. Let ti = min{t : 1 ≤ t ≤ r, i ≤ dt} and ti = r + 1 if i > dr. We also set dr+1 = m. Note that, ti is the smallest t such that Fi ∈/ Ft. We restate this in the following convenient form. Fi ∈ Ft if and only if t < ti. (5.2.6) The following properties will be maintained during the construction of the pseud- oforest decomposition. t (P1) For F ∈ F \ Ft, F contains no cycle, for 1 ≤ t ≤ r. t (P2) |{F ∈ Ft : F covers v}| + |{loops in A }| ≤ m − dt, for 1 ≤ t ≤ r. (P3) |{F ∈ F : F covers v}| + |A| − |D| = m. Note that at the very beginning, when F is the pseudoforest decomposition of H, these conditions hold, and when the construction terminates, |A| = |D| = 0 hence we obtain the desired decomposition. Suppose that at a step in the construction, a pseudoforest Fi ∈ F is replaced ′ ′ ′ by a pseudoforest Fi and we obtain a new family F = F − Fi + Fi . To preserve the condition (P1) for the new family F ′ we only need to make sure that ′ ti (P1’) (Fi ) contains no cycle. If |D| + |A| = 0 then the collection F is the desired decomposition. So suppose that |D| + |A| > 0. First consider the case when D = ∅. If there is a loop at v then let c to be one of highest grade. By condition (P2), and since |Fσc | = m − dσc , there is an ′ ′ ′ Fi ∈ Fσc that does not cover v. Let Fi = Fi + c, A = A − c and D = D. Since ′ Fi does not cover v, Fi is a pseudoforest. Moreover, the only new cycle created by ′ ′ ti the operation in F is the loop c. Note that σc < ti by (5.2.6), so (Fi ) contains ′ ′ no cycle. Hence (P1’) holds for Fi and therefore (P1) holds for F . We check ′ condition (P2) for F . For t ≤ σc < ti, ′ |{F ∈ Ft : F covers v}| = |{F ∈ Ft : F covers v}| + 1, while |{loops in (A′)t}| = |{loops in At}| − 1, 77 5.2. Graded sparse graphs ′ ′ t thus (P2) holds for F . As for t > σc, we have (A ) = ∅ by the maximality of σc, ′ and |Ft| = m − dt. Thus ′ ′ t |{F ∈ Ft : F covers v}| + |{loops in (A ) }| ≤ m − dt for t > σc, thus (P2) also holds for F ′ in this case. (P3) also holds for F ′,A′,D′ since |{F ∈ F ′ : F covers v}| = |{F ∈ F : F covers v}|+1, |D′| = |D|, and |A′| = |A|−1. If there is no loop at v, let a be any edge in A and Fi be a pseudoforest that ′ ′ does not cover a (Fi exists by condition (P3)). Let Fi = Fi + a, A = A − a and D′ = D. (P1) holds for F ′ since no new cycle is created. (P2) also holds since t |{loops in A }| = 0 and |{F ∈ Ft : F covers v}| ≤ |Ft| = m−dt for 1 ≤ t ≤ r. (P3) holds for F ′,A′,D′ since |{F ∈ F ′ : F covers v}| = |{F ∈ F : F covers v}| + 1, |A′| = |A| − 1 and |D′| = |D|. Now suppose that D is not empty. Take an edge e = xy in D. Suppose that e is subdivided into f = vx, g = vy in A. Suppose also that e is an edge of an Fi in F. Case 1: v is not covered by Fi. ′ ′ ′ ′ ′ Let Fi = Fi − e + f + g, F = F − Fi + Fi , D = D − e and A = A − f − g. ′ti First, we check condition (P1’). Suppose by contradiction that (Fi) contains a cycle C then, since (P1) holds for Fi, the cycle C must contain both f and g and ti ti ≤ min{σf , σg} = σe. Hence C − f − g + e is a cycle in Fi , a contradiction. Therefore (P1) holds for F ′. Condition (P3) also holds for F ′,D′,A′ since |{F ∈ F ′ : F covers v}| = |{F ∈ F : F covers v}|+1, |A′| = |A|−2 and |D′| = |D|−1. If F ′,A′ satisfy condition (P2) then we are done. So suppose that F ′,A′ violate (P2). ′ Since (P2) holds for F,A, there exists t < ti such that |{F ∈ Ft : F covers v}| + ′ t ∗ |{loops in (A ) }| > m−dt. Let t be the maximum t < ti such that this inequality holds. Then the following equality holds. t∗ |{F ∈ Ft∗ : F covers v}| + |{loops in A }| = m − dt∗ , (5.2.7) keeping in mind that the sets of loops in A and that in A′ are the same as f, g are not loops. t∗ We claim that there exists a loop in A of grade strictly less than ti. Suppose ∗ by contradiction that |{loops in At }| = |{loops in Ati }|. We have, ti |{F ∈ Fti : F covers v}| + |{loops of A }| ≤ m − dti . (5.2.8) 78 Chapter 5. Inductive constructions and decompositions ∗ Since Fi ∈ Ft \Fti and Fi does not cover v, we have, ∗ ∗ ∗ dti − dt = |Ft \Fti | > |F ∈ Ft \Fti : F covers v|. (5.2.9) ∗ However, from equalities (5.2.7) and (5.2.8) and the assumption |{loops in At }| = |{loops in Ati }|, we obtain ∗ ∗ ∗ |F ∈ Ft \Fti : F covers v| = |{F ∈ Ft : F covers v}|−|{F ∈ Fti : F covers v}| ≥ dti −dt , a contradiction to (5.2.9). Therefore our claim is true. ∗ t ti Let c be a loop in A \ A of maximum grade. Note that σc < ti and hence ′ ′ ′ Fi ∈ Fσc by (5.2.6). If in Fi − f the connected component containing g has a cycle ′′ ′ ′′ ′ ′′ ′′ ′′ then let Fi = Fi −g+c. Otherwise let Fi = Fi −f +c. Set F = F −Fi +Fi , A = A−f −c and D′′ = D−e. We claim that F ′′,A′′ and D′′ verify the conditions (P1), ′′ ′′ (P2), (P3). In fact, since v is already covered by Fi, |A | = |A| − 1, |D | = |D| − 1, (P3) obviously holds. To check (P1), first consider the case where the connected ′ ′ component containing g in Fi − f has a cycle. Then since Fi is a pseudoforest, the ′ connected component containing v in Fi −g has no cycle. Therefore, the connected ′′ ′ ′′ component of Fi = Fi − g + c containing v has only one cycle c and so Fi is a ′′ ′′ pseudoforest. Recall that σc < ti, we have that Fi satisfies (P1’) and hence F satisfies (P1). Now consider the case where the connected component containing ′ ′′ ′ ′′ g of Fi − f has no cycle. Then Fi = Fi − f + c is a pseudoforest and Fi satisfies ′′ ′′ (P1’) as σc < ti. It remains to check the condition (P2) for F ,A . Again, suppose by contradiction that there exists t such that ′′ ′′ |{F ∈ Ft : F covers v}| + |{loops in A |t}| > m − dt. ′′ ′ ′′t Then since |{F ∈ Ft : F covers v}| = |{F ∈ Ft : F covers v}|, and |{loops in A }| ≤ |{loops in A′t}|, it implies ′ ′ |{F ∈ Ft : F covers v}| + |{loops in A |t}| > m − dt. ∗ ∗ ′′ t By the maximality of t , we must have t ≤ t < ti. However, then |{loops in (A ) }| = ∗ |{loops in At}| − 1 holds, since the loop c belongs to At ⊆ At. Therefore, |{F ∈ ′′ ′′ t t Ft : F covers v}|+|{loops in (A ) }| = |{F ∈ Ft : F covers v}|+|{loops in A }| ≤ ′′ ′′ m − dt, a contradiction. We conclude that (P2) also holds for F ,A . Case 2: v is already covered by Fi. In this case we will replace the edge e = xy in Fi by one of the two edges f = vx, g = vy. We remove e from D and the replacing edge from D. Condition (P2) and (P3) then trivially hold. 79 5.2. Graded sparse graphs Case 2.1: The subdivision is proper. ′ We will choose f or g to replace e so that Fi is a pseudoforest and satisfies (P1’). If in Fi − e, one of x and y belongs to a different connected component than ′ that contains v, without loss of generality, let it be x. Then Fi = Fi − e + f is a pseudoforest and (P1’) holds since no new cycle is created. So suppose that in Fi − e, both x and y belong to the same connected component as v. Then the unique cycle in the connected component of Fi containing v must contain e. ti Therefore both Fi −e+f and Fi −e+g are pseudoforests. If both (Fi −e+f) and ti (Fi − e + g) contain a cycle then x and y are on the same connected component ti ti of (Fi − e) and σe ≥ ti. Therefore, there exists a cycle that contains e in Fi . ti This contradiction to the assumption (P1) implies that either (Fi − e + f) or ti (Fi − e + g) does not contain a cycle. Without loss of generality, suppose that ′ ′ ′ ′ ′ Fi = Fi −e+f has this property. Let F = F −Fi +Fi , D = D−e and A = A−f. Then F ′,A′,D′ satisfy condition (P1), (P2) and (P3). Case 2.2: The subdivision is improper. Suppose that f is a loop. It is worth keeping in mind that g and e have the same ends. ′ ′ ′ ′ If σg = σe then we just put Fi = Fi − e + g, F = F − Fi + Fi A = A − g and D′ = D − e. The conditions (P1), (P2), (P3) obviously hold for F ′,A′,D′. So let us suppose that σg > σe and hence σf = σe. ′ ′ If σe ≥ ti or there is no cycle containing e in F , set Fi = Fi − e + g, A = A − g ′ ′ and D = D − e. Then Fi is a pseudoforest and condition (P2), (P3) obviously ′ ti hold. To check condition (P1’), suppose by contradiction that (Fi ) contains cycle ′ ti C . Then the cycle must contain g since Fi has no cycle. We obtain a cycle C in ′ Fi from C by replacing g with e. However, we have assumed that whether there is no cycle containing e in Fi or ti ≤ σe < σg, so we must have ti ≤ σe < σg. ti However, under this condition, C is a cycle in Fi , a contradiction. It means that ′ ′ ′ ′ (P1’) holds for Fi and so F ,A ,D satisfy (P1), (P2) and (P3). Now consider the case when σe < ti and there is a cycle containing e in Fi. Put ′ ′ ′ ′ ′ Fi = Fi − e + f, F = F − Fi + Fi , A = A − f and D = D − e. Since there is ′ ′ a cycle containing e in Fi, Fi is a pseudoforest. Since the only new cycle in Fi is ′ ti that consists of the loop f, which is of grade σf = σe < ti,(Fi ) contains no cycle, ′ ′ ′ ′ so (P1’) holds for Fi . Therefore F ,A ,D satisfy (P1), (P2) and (P3). 2 80 Chapter 5. Inductive constructions and decompositions 5.2.5 Graded sparse matroids This section is devoted to the study of (m, d)-graded sparse matroids, i.e. matroids whose independent sets are the edge sets of (m, d)-graded sparse graphs. The fact that these sets define a matroid was shown by Lee, Streinu and Theran [79] using the matroid circuit axioms. We will provide submodular functions that induce the (m, d)-graded sparse matroids. This gives an alternative proof for the result of [79], and also determines the rank formula for graded sparse matroids. We then use this rank formula to express an (m, d)-graded sparse matroid as a union of m smaller graded sparse matroids. This matroid decomposition results in a simple proof for Theorem 5.2.3. We first describe a general setting for an (m, d)-graded sparse matroid. We think of the ground set as a subset of the edge set of an r-graded graph on n r,m vertices, Kn , in which every vertex is adjacent to m loops of each grade, every pair of vertices are joined by 2m parallel edges of each grade, and n is a large r,m unspecified integer. Since we work with subsets E of the edges of Kn , it is convenient to reformulate the sparsity condition for such an edge set. We say that E is (m, d)-graded sparse if it is the edge set of an (m, d)-graded-sparse subgraph r,m of Kn . It is not difficult to see that an edge set E is (m, d)-graded sparse if and t t only if for every subset F of E, |F | ≤ m|V (F )| − dt for all 1 ≤ t ≤ r, where F is the set of all edges of F which have grade at least t and V (F ) is the set of all vertices which are incident with edges in F . r,m For a subset F of E(Kn ), let fm,d(F ) = m|V (F )| − dt, (5.2.10) where t is the minimum grade of the elements in F when F =6 ∅, and put fm,d(∅) = 0. It is easy to see that fm,d is nonnegative and nondecreasing for d = (d1, d2, . . . , dr) when 0 ≤ d1 ≤ d2 ≤ · · · ≤ dr ≤ 2m − 1 and we will assume henceforth that this is the case. We also have the submodularity property. Lemma 5.2.16. The function fm,d is an intersecting submodular function on the r,m subsets of E(Kn ). Proof. Let F1,F2 be edge sets with F1 ∩ F2 =6 ∅. Let t1, t2 be the minimum grades of edges in F1,F2 respectively. We can suppose without loss of generality that 81 5.2. Graded sparse graphs t1 ≤ t2. Then t1 is the minimum grade of edges in F1 ∪F2, and the minimum grade of edges in F1 ∩ F2 is at least t2 . We have, fm,d(F1) + fm,d(F2) = m|V (F1)| − dt1 + m|V (F2)| − dt2 ≥ m|V (F1 ∪ F2)| − dt1 + m|V (F1 ∩ F2)| − dt2 ≥ fm,d(F1 ∪ F2) + fm,d(F1 ∩ F2) where the first inequality uses the fact that the function F 7→ |V (F )| is submodular. r,m Theorem 2.4.2 now implies that fm,d induces a matroid on E(Kn ). We denote this matroid by M(m, d). We next show that the independent sets of M(m, d) are the (m, d)-graded sparse edge-sets. Theorem 5.2.17. The edge sets of all (m, d)-graded sparse graphs form the inde- pendent sets of the matroid M(m, d). Proof. Suppose that E is the edge set of an (m, d)-graded sparse graph and ∅= 6 F ⊆ E. Let t be the minimum grade of an edge in F . Then |F | = |F t| ≤ m|V (F )| − dt = fm,d(F ). Thus E is independent in M(m, d). Conversely, suppose that E is independent in M(m, d). We show that the graph (V (E),E) is (m, d)-graded sparse. Let ∅= 6 F ⊆ E and let t be the minimum ′ t′ t grade of an edge in F . Then fm,d(F ) = m|V (F )| − dt. For t ≤ t, |F | = |F | ≤ fm,d(F ) = m|V (F )| − dt ≤ m|V (F )| − dt′ , by the (m, d)-sparseness of F and the ′ t′ t′ definition of fm,d. For t ≥ t, |F | ≤ fm,d(F ) ≤ m|V (F )| − dt′ , where the first inequality follows from the fact that F is independent in M(m, d) and the second inequality follows from the definition of fm,d, since the minimum grade of an edge ′ in F t is at least t′ and the sequence d is nondecreasing. Therefore, (V (E),E) is indeed (m, d)-graded sparse. The theorem follows. We will refer to M(m, d) as the (m, d)-graded sparse matroid. Theorem 2.4.2 and Theorem 5.2.17 determine the rank formula for M(m, d). Corollary 5.2.18. The rank function of M(m, d) is given by s rM(m,d)(E) = min fm,d(Fj) + |F0| : {F0,F1 ...,Fs} partitions E . j=1 X In the remainder of this section we assume that d satisfies the stronger hypoth- esis that 0 ≤ d1 ≤ · · · ≤ dr ≤ m. 82 Chapter 5. Inductive constructions and decompositions i i i For 1 ≤ i ≤ m, we define the r-tuple c = (c1, . . . , cr) by i 1 if i ≤ dt, ct = ( 0 otherwise. m i Note that we have dt = i=1 ct, for all t = 1, . . . , r. We will show that the (m, d)-sparse matroid is the union of the m (1, ci)-graded sparse matroids, i.e. P M(m, d) = M(1, c1) ∨ M(1, c2) ∨ · · · ∨ M(1, cm). We will need the following result. Lemma 5.2.19. Suppose H = (V,F ) is a connected graph. If H is (m, d)-graded sparse and |F | > m|V | − dt+1 for some 1 ≤ t ≤ r then there exists an edge e belonging to a cycle of H with σe ≤ t. Proof. First note that since H is (m, d)-graded sparse, |F | ≤ m|V |, and hence we t+1 must have dt+1 ≥ 1. We also have |F | ≤ m|V | − dt+1 from the sparseness of H. Since |F | > m|V | − dt+1, there exists an edge in F of grade at most t. Suppose that every edge of grade at most t in F is a cut-edge of H. Let F1,...,Fs be the edge sets of the connected components of the subgraph of H induced by F t+1 and s F0 = F \ i=1 Fi. Since H is connected and all edges in F0 are cut-edges, we have |F | = s − 1. Thus 0 S s |F | = |F0| + |Fi| i=1 Xs ≤ |F0| + (m|V (Fi)| − dt+1) i=1 X = (s − 1) + m|V | − dt+1 − (s − 1)dt+1 ≤ m|V | − dt+1. This contradiction to the assumption that |F | > m|V | − dt+1 implies that there is an edge e with σe ≤ t that belongs to a cycle of H. Now we are ready to prove the following main result of this section. Theorem 5.2.20. Suppose 0 ≤ d1 ≤ · · · ≤ dr ≤ m. Then the matroid M(m, d) is the union of the m matroids M(1, ci) for i = 1, . . . , m. Proof. By definition i t i if Ii is independent in M(1, c ) then |Ii | ≤ |V (Ii)| − ct (5.2.11) 83 5.2. Graded sparse graphs for all t = 1, . . . , m. Let M = M(1, c1) ∨ · · · ∨ M(1, cm). We need to show that M = M(m, d). First suppose that I is an independent set of M. Then I = I1 ∪ · · · ∪ Im where i Ii is an independent set in M(1, c ) for each i ∈ {1, . . . , m}, by the definition of matroid union. Then for every t, 1 ≤ t ≤ r, and for every i = 1, . . . , m we have, t i |Ii | ≤ |V (Ii)| − ct by (5.2.11). Therefore, m m t t i |I | = |Ii | ≤ (|V (Ii)| − ct) ≤ m|V (I)| − dt, i=1 i=1 X X for all t = 1, . . . , r. We have a similar inequality for every subset J of I. Hence (V (I),I) is indeed (m, d)-graded sparse, and hence I is independent in M(m, d) by Theorem 5.2.17. Conversely, suppose that I is an independent set of M(m, d). We will show that I is also independent in M. We do this by showing that the rank of I in M is equal to |I|. By Theorem 2.3.1, to prove that rM(I) = |I| it is sufficient to show that rM(1,c1)(F ) + ··· + rM(1,cm)(F ) ≥ |F | holds for every subset F of I. Let F1,...,Fk be the edge sets of the connected components of the subgraph H r,m of Kn induced by F . Then |F | = |F1|+···+|Fk| and rM(1,ci)(F ) = rM(1,ci)(F1)+ ··· + rM(1,ci)(Fk) for all i = 1, . . . m. Hence we may suppose that H is connected. Let t∗ be minimum grade of an edge e which belongs to a cycle C of H, where we set ∗ t = 0 if no such edge exists. Then Lemma 5.2.19 implies that |F | ≤ m|V (F )|−dt∗ , ∗ taking dt∗ = 0 when t = 0. On the other hand, it is not difficult to see that |V (F )| − 1, if 1 ≤ i ≤ dt∗ , rM(1,ci)(F ) = ( |V (F )|, otherwise. Since F is independent in M(m, d), this gives rM(1,c1)(F ) + ··· + rM(1,cm)(F ) = m|V (F )| − dt∗ ≥ |F |. The theorem now follows. An immediate consequence of this result is that the union of two graded sparse matroids is again a graded sparse matroid. 84 Chapter 5. Inductive constructions and decompositions ′ ′ ′ Corollary 5.2.21. Let d = (d1, . . . , dr), d = (d1, . . . , dr) be two r-tuples of in- ′ tegers and m, m be positive integers such that 0 ≤ d1 ≤ · · · ≤ dr ≤ m and ′ ′ ′ 0 ≤ d1 ≤ · · · ≤ dr ≤ m . Then the union of two graded sparse matroids M(m, d) and M(m′, d′) is the graded sparse matroid M(m + m′, d + d′), where d + d′ = ′ ′ (d1 + d1, . . . , dr + dr), i.e., M(m + m′, d + d′) = M(m, d) ∨ M(m′, d′). The decomposition of graded sparse graphs is also a consequence of this matroid decomposition result. Proof of Theorem 5.2.3 from matroid decomposition Let m, d and ci, i = 1, . . . , m, be defined as above. It is not difficult to see that a i base Fi of M(1, c ) induces a pseudoforest such that the restriction of Fi on Et is i i a forest if and only if ct = 1. However, by the definition of c , this is equivalent to the condition that the restriction of Fi on Et is a forest if and only if 0 ≤ i ≤ dt. i Moreover, Fi is a spanning forest if and only if c1 > 0 and Fi is a tight spanning i pseudoforest if and only if c1 = 0. Therefore, the matroid decomposition in Theo- rem 5.2.17 induces our desired graph decomposition. 2 85 5.3. (b, l)-sparse graphs 5.3 (b, l)-sparse graphs Let V be a finite set, b : V → Z+, we use bmin to denote min {b(v): v ∈ V }. In this section we suppose that l is an integer satisfying 0 ≤ l < 2bmin. Recall from Chapter 2 that a graph G = (V,E) on V is said to be (b, l)-sparse if i(X) ≤ b(X) − l for every X ⊆ V with i(X) > 0. Contribution: This section provides an inductive construction of (b, l)-sparse graphs for 0 ≤ l ≤ bmin. We also extend the pebble game in [78] to characterize (b, l)-sparse graphs for 0 ≤ l < 2bmin. 5.3.1 Inductive construction of (b, l)-sparse graphs, 0 ≤ l ≤ bmin Throughout this subsection we suppose that 0 ≤ l ≤ bmin. The sparsity condition is then written simply as i(X) ≤ b(X)−l for every ∅= 6 X ⊆ V . Due to the matroidal property of the family of (b, l)-sparse graphs, a (b, l)-sparse graph can always be obtained from a (b, l)-tight graph by deleting some edges. So in this subsection we are interested only in the inductive construction of (b, l)-tight sparse graphs. The proof of our inductive construction for (b, l)-tight sparse graphs G = (V,E) is proceeded by induction on the number of vertices and follows the same line as that for (m, d)-graded tight graphs: first we choose a vertex v ∈ V with some requirement on the number of incident edges, then we show that for this vertex there is a reduction that results in a (b, l)-tight graph on V − v. This proof is relatively simpler than the one for (m, d)- tight graphs but we include it for the sake of completeness. Also, since situations in the two proofs are quite similar, we will abuse several terms already used in the previous section. The required condition for the chosen vertex v is that d(v) = b(v) + k + i(v) with 0 ≤ k + i(v) ≤ b(v). The existence of a vertex satisfying this condition is assured by the following lemma. Lemma 5.3.1. Let G = (V,E) be a (b, l)-tight sparse graph on at least two vertices. Then there exists a vertex v ∈ V such that d(v) = b(v) + k + i(v) for some k ≥ 0 such that 0 ≤ k + i(v) ≤ b(v). Proof. First, for every v ∈ V , we have d(v) − i(v) = i(V ) − i(V − v) ≥ b(V ) − l − (b(V − v) − l) = b(v). 86 Chapter 5. Inductive constructions and decompositions Suppose by contradiction that there is no vertex v ∈ V satisfying the condition in the lemma, then for every v ∈ V , d(v) > 2b(v) holds. Then 2|E| = d(v) > 2 b(v) = 2b(V ) = 2(|E| + l), v∈V v∈V X X a contradiction. The lemma follows. A reduction operation consists of splitting offs and loop deletions. A splitting off at a vertex v of a graph G is an operation that deletes two edge f = vx, g = vy incident to v and adds an edge e = xy. If f is a loop this operation simply deletes f, we refer to this as a loop deletion. If neither f nor g is a loop, we call the operation a proper splitting off. Let v be a vertex of a (b, l)-tight graph G incident with c loops and d(v) = b(v) + k + c, for some 0 ≤ k ≤ m.A k-reduction at v first does c loop deletions at v following by k proper splitting offs at v, then deletes all the remaining b(v) − k edges. Note that when k = c = 0, the k-reduction simply deletes from G the vertex v and all its incident edges. A k-reduction at v is admissible if the resulting graph is also (b, l)-sparse. Suppose that H is a (b, l)-sparse graph and v is a vertex of H with d(v) = b(v) + k + c − 2s, i(v) = max {0, c − s}, (5.3.1) c ≥ 0, k ≥ 1 and 0 ≤ s < k + c ≤ b(v). We can imagine that c is the number of loops at v in G and H is obtained from G after s splitting offs. This condition implies that The number of edges incident to v is strictly greater than b(v) − s. (5.3.2) In fact, this number is d(v) − i(v) = b(v) + k + c − 2s − i(v). If i(v) = 0 then b(v) + k + c − 2s − i(v) = b(v) + k + c − 2s > b(v) − s since s < k + c by assumption 5.3.1. If i(v) > 0 then i(v) = c − s by assumption 5.3.1. Hence b(v) + k + c − 2s − i(v) = b(v) + k + c − 2s − (c − s) = b(v) + k − s > b(v) − s since k ≥ 1 by assumption 5.3.1. Assumption 5.3.1 also implies that b(v) − i(v) ≥ b(v) − (c − s) ≥ k + s ≥ 1. Therefore, we have there are non loop edges at v. (5.3.3) We also say that H is (v, s)-good if i(X) ≤ b(X) − l for every X ⊆ V and i(X) ≤ b(X) − l − s for every X ⊆ V that properly contains v. In the rest of this section we will assume that H is (v, s)-good. (5.3.4) 87 5.3. (b, l)-sparse graphs We will show that we can find a splitting off at v that results in a (v, s + 1)-good graph. Such a splitting off is called feasible. The two potential obstacles for a feasible splitting off are critical sets and crucial sets. A critical set is a set X ⊆ V − v with i(X) = b(X) − l.A crucial set is a set {v} ( X ⊂ V with i(X) = b(X) − l − s. More precisely, we have the following. Lemma 5.3.2. A splitting off of f = vx, g = vy at v is feasible if and only if 1. no critical set contains both x and y, and 2. all crucial sets intersect the set {x, y}. Proof. Let H′ be the graph obtained from H by splitting off f, g at v. First suppose that the splitting off of f, g at v is feasible, i.e., H′ is (v, s + 1)-good. Then for every critical set X ⊆ V − v, b(X) − l = iH (X) ≤ iH′ (X) ≤ b(X) − l. Hence, equality must holds everywhere, in particular, iH (X) = iH′ (X), which implies that X does not contain both x and y. Every crucial set Y must intersect {x, y} since, otherwise, we have b(Y )−l−s−1 ≥ iH′ (Y ) = iH (Y ) = b(Y )−l−s, a contradiction. Conversely, suppose that there is no critical set containing both x and y and all crucial sets intersect {x, y}. Let X be a subset of V − v. If X is not a critical set then iH′ (X) ≤ iH (X) + 1 ≤ b(X) − l holds. If X is a critical set then X does not contain both x and y, thus iH′ (X) = iH (X) = b(X) − l holds. Now let Y be a subset of V that properly contains v. If Y is not crucial then iH′ (Y ) ≤ iH (Y ) ≤ b(Y ) − l − s − 1 holds. If Y is a crucial set then Y intersects {x, y}, so iH′ (Y ) ≤ iH (Y ) − 1 ≤ b(Y ) − l − s − 1 holds. Therefore, splitting off f, g at v is feasible. In the following, we will investigate properties of critical sets and crucial sets. The following relations will be often of use. i(X) + i(Y ) + i(X \ Y,Y \ X) = i(X ∪ Y ) + i(X ∩ (Y ). In particular, i(X) + i(Y ) ≤ i(X ∪ Y ) + i(X ∩ (Y ). It is also useful keeping in mind that b(X) + b(Y ) = b(X ∪ Y ) + b(X ∩ Y ). The first result shows that the intersection and union of two intersecting critical sets are critical sets. 88 Chapter 5. Inductive constructions and decompositions Lemma 5.3.3. Let X,Y be two critical sets with X ∩ Y =6 ∅. Then X ∪ Y and X ∩ Y are critical sets. Moreover, i(X \ Y,Y \ X) = 0. Proof. We have b(X) − l + b(Y ) − l + i(X \ Y,Y \ X) = i(X) + i(Y ) + i(X \ Y,Y \ X) = i(X ∪ Y ) + i(X ∩ Y ) ≤ b(X ∪ Y ) − l + b(X ∩ Y ) − l = b(X) − l + b(Y ) − l. Therefore equality must hold everywhere, so i(X ∪ Y ) = b(X ∪ Y ) − l, i(X ∩ Y ) = b(X ∩ Y ) − l and i(X \ Y,Y \ X) = 0. The lemma follows. An immediate corollary of this lemma is the following. Corollary 5.3.4. The union of a collection of critical sets containing a common element x is a critical set. 2 The next result indicates that the number of edges from a critical set to v is upperbounded by b(v) − s − i(v). Lemma 5.3.5. Let X be a critical set. Then i(v, X + v) ≤ b(v) − s. Proof. Since X is critical we have, b(X)+b(v)−l−s = b(X +v)−l−s ≥ i(X +v) = i(X) + i(v, X + v) = b(X) − l + i(v, X + v). Hence i(v, X + v) ≤ b(v) − s. The following result shows that if two crucial sets intersect in at least two vertices then their union and intersection are also crucial. Lemma 5.3.6. Let X,Y be crucial sets with |X ∩ Y | ≥ 2, then X ∪ Y and X ∩ Y are crucial. Proof. The proof is similar to that about critical sets. We include it for the sake of completeness. b(X) − l − s + b(Y ) − l − s ≤ i(X) + i(Y ) ≤ i(X ∪ Y ) + i(X ∩ Y ) ≤ b(X ∪ Y ) − l − s + b(X ∩ Y ) − l − s = b(X) − l − s + b(Y ) − l − s. Hence equality must hold everywhere, so i(X ∪Y ) = b(X ∪Y )−l−s and i(X ∩Y ) = b(X ∩ Y ) − l − s. The lemma follows. 2 In fact, the following stronger assertion holds: Let X1,...,Xt be critical sets such that Xi ∩ ( 1≤j≤i−1 Xj ) 6= ∅ for every i = 1, . . . , t. Then 1≤i≤t Xi is a critical set. S S 89 5.3. (b, l)-sparse graphs In contrast to critical sets, the number of edges from a crucial set to v is lowerbounded by b(v) − s − i(v) as showed in the following result. Lemma 5.3.7. Let X be a crucial set. Then i(v, X) ≥ b(v) − s. Moreover, i(v, X − v) ≥ 1. Proof. We have b(X) − l − s = i(X) = i(X − v) + i(v, X) ≤ b(X − v) − l + i(v, X) = b(X) − b(v) − l + i(v, X). Therefore i(v, X) ≥ b(v) − s. Now suppose by contradiction that i(v, X − v) = 0. Then i(v) = i(v, X) ≥ b(v) − s > 0 by assumption 5.3.1. Therefore, again by assumption 5.3.1, i(v) = c − s. Hence we have, c − s ≥ b(v) − s, which implies c ≥ b(v), a contradiction to the assumption b(v) ≥ k + c and k ≥ 1. The next result shows that if two crucial sets intersect at only v, then their union contains all the neighbors of v. Lemma 5.3.8. Let X,Y be crucial sets with X ∩ Y = {v}. Then E(v, V ) = E(v, X ∪ Y ). Proof. We have b(v) + k + c − 2s = d(v) (by assumption 5.3.1) ≥ i(v, X) + i(v, Y ) (since X ∩ Y = {v}) = b(v) − s + b(v) − s (by Lemma 5.3.7) ≥ b(v) − 2s + k + c (by assumption 5.3.1) Therefore equality must hold everywhere, in particular, d(v) = i(v, X) + i(v, Y ). Thus E(v, V ) = E(v, X ∪ Y ). We will also need the following result. Lemma 5.3.9. There are at most two different (inclusionwise) minimal crucial sets in H. 90 Chapter 5. Inductive constructions and decompositions Proof. Suppose that X,Y,Z are three minimal crucial sets in H. If |X ∩ Y | ≥ 2 then, by Lemma 5.3.6, X ∩ Y is a crucial set, so X must coincide to Y by the minimality. So let us suppose that X ∩ Y = {v}. Lemma 5.3.8 implies that X ∪ Y contains all the neighbors of v. By Lemma 5.3.7, Z − v contains a neighbor z of v, so z must belong to X ∪ Y . Assume without loss of generality that z ∈ X. Then |X ∩ Z| ≥ 2, thus X ∩ Z is a crucial set, so X = Z by the minimality. The lemma follows. Lemma 5.3.10. There exists a feasible splitting off at v. Proof. Since all crucial sets contain v, a loop deletion at v is alway feasible by Lemma 5.3.2. So if there exists a loop at v, the lemma holds. Therefore we may suppose that i(v) = 0. Let us first consider the case that there exist two different minimal crucial sets X and Y . Then let f = vx be an edge in E(v, X − v) and g = vy be an edge in E(v, Y − v) (their existence is guaranteed by Lemma 5.3.7). Our aim is to prove that splitting off f, g at v is feasible. Since every crucial set contains either X or Y by Lemma 5.3.9, the second condition on crucial sets of Lemma 5.3.2 is sastisfied. It remains to show that there is no critical set containing both x and y. Suppose on the contrary that U is a critical set containing both x and y. A similar proof to that of Lemma 5.3.3 shows that i(U \ X,X \ U) = 0, contradicting the fact that g = vy ∈ E(U \ X,X \ U). Therefore, both conditions in Lemma 5.3.2 is satisfied, so splitting off f, g at v is feasible. Now let us consider the case that there exists at most one minimal crucial set. If one exists, let X denote the minimal crucial set, otherwise, let X = V . Let f = vx be an edge in E(v, X − v) by Lemma 5.3.7. If there exists a critical set containing x, let U be the maximal one (U is in fact the inclusionwise maximum critical set containing x by Corollary 5.3.4). Then by Lemma 5.3.5, i(v, U + v) ≤ b(v) − s while i(v, V ) > b(v) − s by (5.3.2). Therefore, there exists an edge g = vy ∈ E(v, V − v − U). Splitting off f, g at v is then feasible by Lemma 5.3.2. Now we are ready to prove the main reduction lemma. Lemma 5.3.11 (Reduction lemma). If G = (V,E) is a (b, l)-tight graph then there exists an admissible k-reduction at a vertex v of G. Proof. We choose a vertex v ∈ V incident with c loops and d(v) = b(v) + k + c for some 0 ≤ k ≤ b(v) which exists by Lemma 5.3.1. If k = 0 we simply delete 91 5.3. (b, l)-sparse graphs the vertex v. The obtained graph H is obviously (b, l)-sparse and it is tight as well since |E(H)| = |E(G)| − (d(v) − i(v)) = b(V ) − l − b(v) = b(V (H)) − l. If k ≥ 1, we initialize the induction steps with s = 0. Lemma 5.3.2 shows that, as long as s < k + c, we can find a splitting off at v such that the obtained graph is (v, s + 1)-good. Therefore, we can do k + c feasible splitting offs and then delete the vertex v. The obtained graph H is (b, l)-sparse by the definition of feasible splitting offs. It remains to check that H is tight. Using the tightness of G we have |E(H)| = |E(G)| − c − k − (b(v) − k) = b(V ) − l − c − k − (b(v) − k) = b(V (H)) − l. The lemma follows. Next we define the reverse operation of k-reduction, that we call k-extension. Let b : V → Z+ an integer-valued function on a set V .A k-extension on a (b, l)-sparse graph H with V (H) ⊂ V is an operation that 1. chooses a vertex v ∈ V − V (H); 2. delete k ≤ b(v) edges ei = xiyi, i = 1, . . . k from H; 3. for each i = 1, . . . , k, add edges vxi, vyi to H; then 4. adds b(v) − k edges incident with v such that there are at most b(v) − l loops among them. We call this operation a k-extension on e1, . . . , ek with new vertex v. Note that the number of edges increases by b(v). The main theorem of this section characterizes (b, l)-tight graphs in term of inductive construction. Theorem 5.3.12. Let b : V → Z+ and 0 ≤ l ≤ min {b(u): u ∈ V }. A graph G = (V,E) on V is (b, l)-tight if and only if it can be constructed from a graph on a single vertex v0 ∈ V with b(v0) − l loops by a sequence of k-extensions. Proof. The “if” part follows from Lemma 5.3.11 and by induction. We prove the “only if” part by induction. The graph on a single vertex v0 with b(v0) − l loops is clearly (b, l)-tight. Suppose that a graph G = (V,E) is obtained from a (b, l)-tight graph H by a k-extension on k edges ei = xiyi for i = 1, . . . , k with the vertex v ∈ V − V (H). We check the (b, l)-tightness of G. For a subset X of V − v, iG(X) ≤ iH (X) ≤ b(X) − l by the sparseness of H. If X = {v}, iG(v) ≤ b(v) − l by the condition on the number of added loops. For a subset X of V that contains v, if X contains both ends xi, yi of an edge ei, then when comparing to EH (X), in EG(X) two edges vxi, vyi are added while an edge ei is deleted. So the total 92 Chapter 5. Inductive constructions and decompositions contribution of the deletion of ei and addition of vxi, vyi to EG(X) is at most one. The contribution of any other added edge to EG(X) is also at most one. Therefore, iG(X) ≤ iH (X − v) + b(v) ≤ b(X − v) − l + b(v) = b(X) − l holds. Hence G is (b, l)-sparse. Furthermore, since H is tight, |E(G)| = |E(H)| + b(v) = b(V (H)) − l + b(v) = b(V (G)) − l. The theorem follows. 5.3.2 (b, l)-pebble games In this subsection, we introduce (b, l)-pebble games which generalizes (m, l)-pebble games by Lee and Streinu [78]. We show that (b, l)-sparse graphs can be charac- terized using (b, l)-pebble games. Let V be a finite set of vertices, b : V → Z+ and 0 ≤ l ≤ 2bmin. Let G = (V,E) be an undirected graph on V . In a (b, l)-pebble game, each vertex v of V is provided with b(v) pebbles. We consider an unoriented edge uv of G and try to orient it using the following rules. 1. Add-edge move: If there are totally at least l + 1 pebbles at u and v and p is a pebble at u, then put p on uv and orient uv from u to v. We say that uv is covered by p. 2. If there are totally at most l pebbles at u and v, then try to collect at least l + 1 pebbles at u and v using pebble-slide moves which will be described later. 3. If it is impossible to collect at least l + 1 pebbles at u and v then discard the edge uv. A pebble-slide move concerns an oriented edge uv (oriented from u to v). Let p be the pebble placed on uv and p′ a pebble at v. A pebble-slide move puts p back to u, places p′ on uv and reorients uv from v to u. Note that in a pebble game, a pebble stays either in its original vertex or on an edge adjacent to this vertex and once an edge is oriented, it remains an oriented edge (althought may be reoriented to the inverse direction). At the end of the game, we obtain a subgraph H of G whose edges are all the edges of G that are oriented by the pebble game. We say that H is constructed by a (b, l)-pebble game. The (b, l)-sparseness of H follows from the following lemma. At some stage in the pebble game, for a subset U of V , let peb(U) denote the total number of pebbles at all the vertices of U, span(U) denote the number of oriented edges with both ends in U and out(U) denote the number of oriented edges with tail in U and head in V \ U. If U = {v}, we simply 93 5.3. (b, l)-sparse graphs use peb(v), span(v) and out(v) to refer to peb(U), span(U) and out(U). We denote by D the directed graph obtained at stage of a game. Lemma 5.3.13. During the pebble game, the following conditions always hold on D for every v ∈ V and for every U ⊆ V . 1. peb(v) + span(v) + out(v) = b(v); 2. peb(U) + span(U) + out(U) = b(U); 3. peb(U) + out(U) ≥ l if l ≤ bmin or |U| ≥ 2; 4. span(U) ≤ b(U) − l if l ≤ bmin or |U| ≥ 2. Proof. The first two equalities follows from the fact that in a pebble game, a pebble always stays either in its original vertex or on an edge oriented out from this vertex. Hence peb(v) + span(v) + out(v) is the total number of pebbles originally placed at v and so is equal to b(v). Similarly, peb(U) + span(U) + out(U) is the total number of pebbles originally placed at the vertices of U and so is equal to b(U). The last inequality follows easily from the second equality and the third inequality. Now we prove the third inequality by induction on the number of moves. At the beginning, all the pebbles are placed at the vertices, so peb(U) + out(U) = b(U) ≥ l if l ≤ bmin or |U| ≥ 2. It is sufficient to show that after each move, this condition always holds. First consider an add-edge move that orients an unoriented edge uv from u to v. Then by definition, before the move pev(u) + peb(v) ≥ l + 1, so after the move pev(u)+peb(v) ≥ l holds. Hence if both u, v are in U we have peb(U)+out(U) ≥ l. If at most one of u, v belongs to U then this move does not decrease out(U), and if peb(U) is decreased by this move it is because u ∈ U, v∈ / U and one pebble from u is put on uv, which decreases peb(U) by one. However, in this case, the move increases out(U) by one. Therefore the condition holds after an add-edge move. Now consider a pebble-slide move on an oriented edge uv (oriented from u to v.) If both u, v are in U or both u, v are not in U then the move does not change peb(U) + out(U), so the condition holds after this pebble-slide move. If u ∈ U and v ∈ V \ U then the move decreases out(U) by one in reorienting uv, but it increases peb(U) by one by pushing back a pebble on uv to u. So in this case, peb(U) + out(U) remains the same. If u ∈ V − U and v ∈ U, the move decreases peb(U) by one but increases out(U) by one. So peb(U) + out(U) remains the same in this case too. Therefore, the condition holds after a pebble-slide move. The lemma follows. 94 Chapter 5. Inductive constructions and decompositions An immediate consequence of Lemma 5.3.13 is the following. Corollary 5.3.14. The resulting graph in (b, l)-pebble game is a (b, l)-sparse graph. The next result shows that if we play the (b, l)-pebble game on a (b, l)-sparse graph G then no edge is discarded. Lemma 5.3.15. Let H be the underlying graph of the directed graph obtained at a stage of the (b, l)-pebble game on a graph G. Suppose that an unoriented edge uv is then considered. Then uv is discarded by the pebble game if and only if uv is in the closure of E(H) in the (b, l)-count matroid on G. Proof. If uv is not discarded by the (b, l)-pebble game then it is added to H. so Corollary 5.3.14 implies that H + uv is (b, l)-sparse. Therefore uv is not in the closure of E(H) in the (b, l)-count matroid on G. Now suppose that uv is discarded by the pebble game. Let D be the directed graph obtained at the stage just before we discard uv and after all the possible pebble-slide moves to collect pebbles at u and v. Let Reach(u) and Reach(v) denote the set of vertices reachable from u, v, respectively, in D and let X = Reach(u) ∪ Reach(v). Since no more pebble can be collected to u and v, there is no pebble on every vertex x of X other than u, v, otherwise, we can do pebble-slide moves on the dipath from, say, u to x to collect one more pebble at u and v. Hence peb(X) = peb(u) + peb(v) ≤ l. Moreover, out(X) = 0. Therefore, condition 2 of Lemma 5.3.13 implies that iH (X) = span(X) = b(X)−peb(u)−peb(v) ≥ b(X)−l. Since H is (b, l)-sparse, iH (X) = b(X) −l holds. Hence H + uv is not (b, l)-sparse, so uv is in the closure of E(H) in the (b, l)-count matroid on G. Therefore we obtained the following characterization of (b, l)-sparse graphs in term of pebble games. Theorem 5.3.16. A graph is (b, l)-sparse if and only if it is obtained by a (b, l)- pebble game. 95 5.4. Packing of matroid-based arborescences 5.4 Packing of matroid-based arborescences 5.4.1 Introduction Recall from Section 2.2 that an out-arborescence (resp., in-arborescence) is a di- rected graph in which each vertex has in-degree (resp., out-degree) at most 1 and there is exactly one vertex r with in-degree (resp., out degree) 0. Namely, an out-aborescence is a tree where all edges are oriented away from a root node while an in-arborescence is a tree where all edges are oriented toward a root node. The problem of packing arborescences has important applications in many practical problems such as evacuation, commodity, broadcasting,. . . For example, in evacua- tion situation (tsunami, earthquake, fire,. . . ), an in-arborescence represents roads used by refugees while an out-arborescence can represent the roads use by emer- gency vehicles. Since each road has a limited capacity, it is preferable to have several disjoint paths from each place to the safe place, where comes the necessity of considering packing of arborescences. In broadcast networks or commodity net- works, informations or commodities are also sent along arcs of arborescences. The existence of a packing of arborescences satisfying some required conditions makes sure that the informations or commodities are sent without interference. From a theoretical view point, the following two problems are fundamental. 1. Given a digraph (a network with pre-determined orientation of edges), discern whether there exists a certain number of arc-disjoint arborescences satifying some conditions. 2. Given an undirected graph, discern whether one can orient the edges of the graph such that in the oriented graph there exists a certain number of arc- disjoint arborescences satisfying some conditions. This problem reduces to the problem of discerning whether there exists a certain number of edge- disjoint trees satisfying the corresponding conditions. The earliest fundamental results on these problems are due to Nash-Williams, Tutte and Edmonds. Theorem 5.4.1 (Nash-Williams [82], Tutte [104]). A graph G = (V,E) contains k edge-disjoint spanning trees if and only if the inequality eG(P) ≥ k|P| − k holds for every partition P of V . 96 Chapter 5. Inductive constructions and decompositions Theorem 5.4.1 is the dual of Theorem 5.1.1 Theorem 5.4.2 (Edmonds [31]). Let D = (V,A) a digraph and r ∈ V . The digraph D contains k arc-disjoint spanning out-arborescences rooted at r if and only if ρD(X) ≥ k for every nonemty X ⊆ V − r. Frank pointed out that the undirected result of Nash-Williams and Tutte can be obtained easily from its directed counterpart of Edmonds through an orientation result. Theorem 5.4.3 (Frank [39]). Let G = (V,E) be an undirected graph and r a vertex of G. There exists an orientation D of G such that ρD(X) ≥ k for every nonempty X ⊆ V − r if and only if eG(P) ≥ k|P| − k for every partition P of V . The above result of Edmonds is generalized in many ways, among them we can name results on packing of branchings, packing of Steiner arborescences3. One of the notable generalizations is the following by Kamiyama, Katoh and Takizawa. For a digraph D = (V,A) and a vertex r ∈ V , let Reach(r) denote the set of all vertices v reachable from r, i.e., there exists a directed path from r to v in D. In the remainder of this section we will simply use “arborescence” to refer to “out-arborescence”. Theorem 5.4.4 (Kamiyama, Katoh and Takizawa [71]). Let D = (V,A) be a digraph and R = {r1, . . . , rt} a multiset of elements in V . There exist k arc-disjoint arborescences T1,...,Tk in D such that Ti is rooted at ri and spans Reach(ri) if and only if ρD(X) ≥ p(X) for every nonempty X ⊂ V, where p(X) denotes the number of i ∈ {1, . . . , t} such that ri ∈/ X and there exists a directed path from ri to an element in X. Fujishige generalizes this results for packing arc-disjoint arborescences spanning convex sets. A convex set in a digraph D = (V,A) is a set U ⊆ V such that every dipath from u to v with u, v ∈ U lies completely in U. Note that, for each i ∈ {1, . . . , t}, Reach(ri) is obviously a convex set. 3The problem of discerning whether a directed graph contains k edge-disjoint Steiner arbores- cences is NP-hard in general, but it is polynomial when some condition on the in-degree of terminals set is satisfied [11] 97 5.4. Packing of matroid-based arborescences Theorem 5.4.5 (Fujishige [41]). Let D = (V,A) be a digraph, R = {r1, . . . , rt} a multiset of elements of V and U1,...,Ut convex sets with ri ∈ Ui for i = 1, . . . , t. There exist arc-disjoint arborescences T1,...,Tt rooted at r1, . . . , rt such that Ti spans Ui, for i = 1, . . . , t, if and only if (U1,...,Uk) ρD(X) ≥ pR (X) for all nonempty X ⊂ V (U1,...,Uk) where pR (X) denotes the number of i ∈ {1, . . . , t} such that Ui ∩ X =6 ∅ and ri ∈/ X. We generalize the result of Edmonds in another direction. Suppose that S = {s1,..., st} is a finite set and D = (V,A) is a digraph and π : S → V is a map. We can think of π as a placement of elements of S on vertices of D. An S-rooted arborescence of D is a pair (T, s) where T is an arborescence of D rooted at a vertex r ∈ V with π(s) = r. We will also call s the root of T , S the root set of (D, π) and we say that T is rooted at s. When S is clear from the context we simply call an S-rooted arborescence a rooted-arborescence. Let M be a matroid on S. We call the quadruple (D, M, S, π) a matroid-based rooted-digraph.A matroid-based packing of arborescences is a set {(T1, s1),..., (Tt, st)} of pairwise arc-disjoint S-rooted arborescences of D such that for each v ∈ V , the −1 set {si ∈ S : v ∈ V (Ti)} forms a base of M (Figure 5.1). We denote SX = π (X) for X ⊆ V and Sv = S{v} for v ∈ V . Our main result in this section is the following. Theorem 5.4.6. A matroid-based rooted-digraph (D, M, S, π) has a matroid-based packing of arborescences if and only if the following conditions are satisfied. 1. Sv is independent for every v ∈ V . 2. For every nonempty subset X ⊂ V , the inequality ρD(X) ≥ rM(S) − rM(SX ) holds. A map π satisfying condition 1 in Theorem 5.4.6 is said to be M-independent. A quadruple (D, M, S, π) satisfying condition 2 in this theorem is said to be rooted- connected. Our result is in fact motivated by the study of the rigidity of frameworks, or more precisely, by the work of Katoh and Tanigawa on matroid-based packing of trees, which takes it motivation from the study on infinitesimal rigidity of frame- works with boundaries (see Section 6.2). Let us take a brief look at their result. 98 Chapter 5. Inductive constructions and decompositions s π( 1)• • π(s2) T1 • T2 T3 π(s3)• • Figure 5.1: A matroid-based packing of rooted-arborescences where the set of the inde- S pendent sets of the matroid on S = {s1, s2, s3} is 2 \ S. Let G = (V,E) be a graph, M a matroid on a finite set S = {s1,..., st} and π : S → V a map. The quadruple (G, M, S, π) is then called a matroid-based rooted-graph. An S-rooted tree of G is a pair (T, s) where T is a tree of G and s ∈ S with π(s) ∈ V (T ). The element s is then called the root of (T, s) and the set S is called the root set of G. A matroid-based packing of rooted-trees of (G, M, S, π) is a set {(T1, s1),..., (Tt, st)} of pairwise edge-disjoint S-rooted trees of G such that for each v ∈ V , the set {si : v ∈ V (Ti)} forms a base of M. The following undirected counterpart of Theorem 5.4.6 is slightly stronger than that stated in [73] but in fact is implicit in [73]. Theorem 5.4.7 (Katoh and Tanigawa [73]). Let (G, M, S, π) be a matroid-based rooted graph. There exists a matroid-based packing of rooted-trees in (G, M, S, π) if and only if the following conditions are satisfied. 1. Sv is independent in M for each v ∈ V , and 2. eG(P) ≥ rM(S)|P| − X∈P rM(SX ), for all partition P of V . P A map π satisfying condition 1 in Theorem 5.4.6 is said to be M-independent.A quadruple (G, M, S, π) satisfying condition 2 in this theorem is said to be partition- connected. s π( 1)• • π(s2) T1 • T2 T3 π(s3)• • Figure 5.2: A matroid-based packing of rooted-trees where the set of the independent S sets of the matroid on S = {s1, s2, s3} is 2 \ S. 99 5.4. Packing of matroid-based arborescences In the same manner as Theorem 5.4.1 is obtained from its directed counter part, Theorem 5.4.2, Theorem 5.4.7 follows from our Theorem 5.4.6 through the following general orientation result of Frank. V Theorem 5.4.8 (Frank [38]). Let G = (V,E) be a graph and h : 2 → Z+ an intersecting supermodular non-negative non-increasing set-function. There exists an orientation D of G such that ρD(X) ≥ h(X) for all non-empty X ⊂ V if and only if for every partition P of V , eG(P) ≥ h(X). X∈P X Proof of Theorem 5.4.7 from Theorem 5.4.6 First suppose that there exists a matroid-based packing {(T1, s1),..., (Tt, st)} of rooted-trees in (G, M, S, π). Let D be an orientation of G where each rooted- ′ ′ ′ tree (Ti, si) becomes a rooted-arborescence (Ti , si). Then {(T1, s1),..., (Tt , st)} is a matroid-based packing of rooted-arborescences in (D, M, S, π). By Theorem 5.4.6, π is M-independent and (D, M, S, π) is rooted-connected and hence, by Theorem 5.4.8,(G, M, S, π) is partition-connected. Now suppose that π is M-independent and (G, M, S, π) is partition-connected. By Theorem 5.4.8, there exists an orientation D of G such that (D, M, S, π) is rooted-connected. Then, by Theorem 5.4.6, applied to the function h(X) = rM(S) − rM(SX ), there exists a matroid-based packing of rooted-arborescences in (D, M, S, π) which provides, by forgetting the orientation, a matroid-based pack- ing of rooted-trees in (G, M, S, π). 2 In fact, Katoh and Tanigawa deduced Theorem 5.4.7 from its dual form given below. We show that Theorem 5.4.7 also implies Theorem 5.4.9. Theorem 5.4.9 (Katoh and Tanigawa [73]). Let (G, M, S, π) be a matroid-based rooted-graph. Let M be of rank k with rank function rM. Then (G, M, S, π) admits a matroid-based rooted-tree decomposition if and only if π is M-independent, |E|+ |S| = k|V | and |E(X)| ≤ k|X| − k + rM(SX ) − |SX | for all non-empty X ⊆ V . Proof. We first prove the necessity. The M-independence of π is obviously neces- sary since each vertex is always covered by the trees rooted at elements s placed at that vertex. Let D be the orientation of G where each rooted-tree in the de- composition is oriented out from its root. Then in each arborescence, each vertex 100 Chapter 5. Inductive constructions and decompositions has in-degree 1 except to the root which has in-degree 0. Moreover, each vertex is covered by exactly k arc-disjoint arborescences, and there are |S| arborescences in total. So the total number of in-degrees of all vertices in D is k|V |−|S| which is also |E|. Therefore |E| + |S| = k|V | holds. Now for each vertex set X ⊆ V , the total number of in-degrees of all vertices in X is v∈X ρD(v) = k|X| − |SX |. Moreover, a vertex v in X can not be covered by more than r (S ) arborescences rooted P M X in vertices of X, thus v must be covered by at least k − rM(SX ) arborescences rooted at vertices in V \ X. Hence, there are at least k − rM(SX ) arcs entering X. Therefore |E(X)| = ρD(v) − ρD(X) ≤ k|X| − k + rM(SX ) − |SX | v∈X X holds. Now suppose that the conditions hold. For every partition P of V , by the inequality applied for X ∈ P and by |E| + |S| = k|V |, we have eG(P) = |E| − |E(X)| X X∈P ≥ |E| − (k|X| − k + rM(SX ) − |SX |) X X∈P = k|P| − rM(SX ). X∈P X Hence (G, M, S, π) is partition-connected. Then, since π is M-independent, Theo- rem 5.4.7 implies that (G, M, S, π) admits a matroid-based packing of rooted-trees which, by |E| + |S| = k|V |, must be a matroid-based rooted-tree decomposition of (G, M, S, π). Contribution and organization: In Section 5.4.2 we provide a proof of our main result in this chapter, Theorem 5.4.6. This proof is short and relatively simple. Therefore, as discussed above, we obtain a short alternative proof for Theorem 5.4.7 and hence Theorem 5.4.9 of Katoh and Tanigawa. In Section 5.4.3 we offer a polyhedral description for the matroid-based packing of rooted-arborescences. Section 5.4.4 considers the algorithmic aspect of our packing problem. We show that the problem of discerning the existence of a matroid-based packing of rooted- arborescences is in P and the corresponding optimization problem also can be solved in polynomial time. Lastly, Section 5.4.5 discusses related problems and a recently extended result based on our result. 101 5.4. Packing of matroid-based arborescences Results in this section are from a joint work with Olivier Durand de Gevigney and Zolt´an Szigeti [28]. 5.4.2 Proof of the main theorem First we prove the necessity of the conditions. Proof of necessity in Theorem 5.4.6 Suppose that there exists a matroid-based packing T = {(T1, s1),..., (Tt, st)} of rooted-arborescences in (D, M, S, π). Let v be an arbitrary vertex of V and X a vertex set containing v. Since the root set of all rooted-arborescences covering v is a base of M, the root set of all rooted-arborescences of T covering v with root in X is an independent set, in particular, Sv is independent. Hence, there are at most rM(X) arborescences of T with root in X that cover v. Therefore, there are at least rM(S) − rM(X) rooted-arborescences of T covering v which has root in V \ X. In each of these rooted-arborescences, there is at least one arc entering X. Since these rooted-arborescences are arc-disjoint, the number of arc entering X is at least rM(S) − rM(X), that is (D, M, S, π) is rooted-connected. 2 To prove the sufficiency, let us introduce the following definitions. A ver- tex set X is called tight if ρD(X) = rM(S) − rM(SX ). For vertex sets X and Y , we say that Y dominates X if SX ⊆ clM(SY ). Note that since, for Q ⊆ S, clM(clM(Q)) = clM(Q), domination is a transitive relation. We say that an arc uv is bad if v dominates u, otherwise it is good. Note that only good arcs are potential candidates for arcs of rooted-arborescences in a matroid-based packing of rooted-arborescences. Claim 1. Suppose that (D, M, S, π) is rooted-connected. Let X be a tight set and v a vertex of X. (a) If Y is a tight set that contains v, then X ∩Y and X ∪Y are tight. Moreover, if s ∈ clM(SX ) ∩ clM(SY ), then s ∈ clM(SX∩Y ). (b) If no good arc exists in D[X], then v dominates X. 102 Chapter 5. Inductive constructions and decompositions Proof. (a) We have ρD(X) + ρD(Y ) = rM(S) − rM(SX ) + rM(S) − rM(SY ) ≤ rM(S) − rM(SX ∪ SY ) + rM(S) − rM(SX ∩ SY ) ≤ rM(S) − rM(SX∪Y ) + rM(S) − rM(SX∩Y ) ≤ ρD(X ∪ Y ) + ρD(X ∩ Y ) ≤ ρD(X) + ρD(Y ). The first equility holds by the tightness of X and Y , the second inequality holds by the submodularity of rM, the fourth holds by the rooted connectedness of (D, M, S, π) and the last holds by the submodularity of ρD. Therefore we have equality everywhere and hence X ∪ Y , X ∩ Y are tight. We also have rM(SX ) + rM(SY ) = rM(SX ∪ SY ) + rM(SX ∩ SY ) holds. If s ∈ clM(SX ) ∩ clM(SY ) then rM(SX∪Y ) + rM(SX∩Y ) = rM(SX ) + rM(SY ) = rM(SX ∪ s) + rM(SY ∪ s) ≥ rM(SX ∪ SY ∪ s) + rM((SX ∩ SY ) ∪ s) ≥ rM(SX∪Y ∪ s) + rM(SX∩Y ∪ s) ≥ rM(SX∪Y ) + rM(SX∩Y ). Therefore, again, equality holds everywhere, in particular, rM(SX∩Y ∪s) = rM(SX∩Y ) holds, which means that s ∈ clM(SX∩Y ). (b) Let us denote by Y the set of vertices from which v is reachable in D[X]. We show that v dominates Y and Y dominates X and then, since domination is transitive, (b) follows. For all y ∈ Y , there exists a directed path y = vl, . . . , v1 = v from y to v in D[X]. Since no good arc exists in D[X], Sy = Svl ⊆ clM(Svl−1 ) ⊆ · · · ⊆ clM(Sv1 ) = clM(Sv). Hence SY = y∈Y Sy ⊆ clM(Sv) and v dominates Y. By the definition of Y , every arc of D that enters Y enters X as well. Then, by S the rooted-connectedness of (D, M, S, π), the tightness of X and the monotonicity of rM, we have rM(S) − rM(SY ) ≤ ρD(Y ) ≤ ρD(X) = rM(S) − rM(SX ) ≤ rM(S) − rM(SY ). Thus equality holds everywhere and Y dominates X. Now we can prove the main result. 103 5.4. Packing of matroid-based arborescences Proof of sufficiency in Theorem 5.4.6 We prove by induction on the number of good arcs. First consider the case when no good arc exists. Let Ti be the arborescence containing only on vertex ri = π(si) for i = 1, . . . , t. Then {(Ti, si): i = 1, . . . , t} forms a matroid-based packing of rooted-arborescences in (D, M, S, π). To prove this it is sufficient to show that Sv is a base of M for all v ∈ V . Indeed, let Xv be the set of vertices from which v is reachable in D. Then ρD(Xv) = 0 holds. The rooted- connectedness of (D, M, S, π) implies that rM(S) − rM(SX ) = 0. However, since there are only bad arc in (D, M, S, π), rM(SXv ) = rM(Sv). Therefore rM(Sv) = rM(S) holds. Combined with the M-independence of π we obtain that Sv is a base of M. Now suppose that at least one good arc exists. For a good arc uv ∈ A and s ∈ Su \cl(Sv) we define a new matroid-based rooted digraph as follows. We set D′ = D − uv, S′ = S ∪ s′ where s′ is a new element. We extend M to a matroid M′ on S′ by defining s′ as an element parallel to s. Lastly, we obtain a placement π′ of S′ in V from π by placing the new element s′ at v. ••• ••• ••• ••• π(s′) v v • π(s) • π(s) u u in D in D′ Figure 5.3: Changing rooted-arborescences. A matroid-based packing of arborescences T ′ of (D′, M′, S′, π′) will provide a matroid-based packing of arborescences T for (D, M, S, π) as follows. Let (T ′′, s) = (T ∪ T ′ ∪ uv, s), T = T ′ ∪ {(T ′′, s)}) \{(T, s), (T ′, s′)}. Since s and s′ are parallel in M′, the rooted-arborescences (T, s) and (T ′, s′) of P′ are vertex disjoint, so (T ′′, s) is a rooted-arborescence. Then T ′ is a matroid- based packing of rooted-arborescences in (D, M, S, π). 104 Chapter 5. Inductive constructions and decompositions Since D′ has less good arcs than D if π′ is M′-independent and (D′, M′, S′, π′) is rooted-connected then (D′, M′, S′, π′) has a matroid-based packing of arbores- cences by induction. Since the M′-independence of π′ is trivial from the M- independence of π and the fact that s ∈ Su − cl(Sv), for the existence of a matroid- based packing of arborescences in (D′, M′, S′, π′), by induction hypothesis it is ′ ′ ′ ′ sufficient to find a good arc uv and s ∈ Su \ cl(Sv) such that (D , M , S , π ) is rooted-connected. Assume that such a good arc does not exist. Let uv ∈ A be a good arc ′ ′ ′ ′ and s ∈ Su \ cl(Sv). Then since (D , M , S , π ) is not rooted-connected, there ′ ′ ′ exists ∅= 6 Xs ⊂ V such that ρD (Xs) < rM(S) − rM (SXs ). Hence, by the rooted- connectedness of (D, M, S, π) and the monotonicity of rM′ , ρD′ (Xs) + 1 ≥ ρD(Xs) ≥ rM(S) − rM(SXs ) ′ ′ ≥ rM(S) − rM (SXs ) ≥ ρD′ (Xs) + 1. So equality holds everywhere and thus uv enters X, Xs is tight in (D, M, S, π) and s ∈ clM(SXs ). Hence, by Claim 1(a), X = ∪s∈Su\cl(Sv )Xs is tight and, by v ∈ X, Su = (Su \ cl(Sv)) ∪ (Su ∩ cl(Sv)) ⊆ cl(SX ) ∪ cl(SX ) = cl(SX ). So we have proved that every good arc uv enters a tight set X that dominates u. (5.4.1) Among all pairs (uv, X) satisfying (5.4.1) choose one with X minimal. Since X dominates u but v does not dominate u, v does not dominate X. Then, by Claim 1(b), there exists a good arc u′v′ in D[X]. Then, by (5.4.1), u′v′ enters a tight set Y that dominates u′. By v′ ∈ X ∩ Y , the tightness of X and Y , u′ ∈ X, Su′ ⊆ clM(SY ) and Claim 1(a), we have that X ∩ Y is tight and Su′ ⊆ clM(SX∩Y ). Since the good arc u′v′ enters the tight set X ∩ Y that dominates u′ and X ∩ Y is a proper subset of X (since u′ ∈ X \ Y ), this contradicts the minimality of X. 2 5.4.3 Polyhedral description In this section we provide a polyhedron describing the matroid-based packings of rooted-arborescences. We need the following general result of Frank [37]. 105 5.4. Packing of matroid-based arborescences V Theorem 5.4.10 (Frank [37]). Let D = (V,A) be a digraph, h : 2 → Z+ a non- negative intersecting supermodular set-function such that ρD(X) ≥ h(X) for every X ⊆ V . Then the polyhedron defined by the following linear system is integer: 1 ≥ x(a) ≥ 0 for all a ∈ A, − x(RD(X)) ≥ h(X) for all non-empty X ⊆ V. The following theorem is a corollary of Theorems 5.4.6 and 5.4.10. Theorem 5.4.11. Let (D = (V,A), M, S, π) be a matroid-based rooted-digraph where M is of rank k with rank function rM and suppose that π is M-independent. The convex hull PM,D of characteristic vectors of the arc sets of the matroid-based packing of rooted-arborescences in (D, M, S, π) is given by the system 1 ≥ x(a) ≥ 0 for all a ∈ A, (5.4.2) − x(RD(X)) ≥ k − rM(SX ) for all non-empty X ⊆ V, (5.4.3) x(A) = k|V | − |S|. (5.4.4) In particular, there exists a matroid-based packing of rooted-arborescences in (D, M, S, π) if and only if the polyhedron PM,D is not empty. Proof. By Theorem 5.4.10, the polyhedron P determined by the subsystem (5.4.2) and (5.4.3) is integer. For every x in this polyhedron, we have − x(A) = x(RD(v)) ≥ (k − rM(Sv)) ≥ (k − |Sv|) = k|V | − |S|. (5.4.5) v V v V v V X∈ X∈ X∈ Therefore, (5.4.4) is a valid inequality for this polyhedron P and hence PM,D is a face of P . Thus PM,D is integer. A characteristic vector of a matroid-based packing of rooted-arborescences in (D, M, S, π) obviously belongs to PM,D. Conversely, if ′ x is a vertex of PM,D then it is integer. Let A = {a ∈ A : x(a) = 1} and D′ = (V,A′). Then the matroid-based rooted digraph (D′, M, S, π) is rooted- connected. Together with the assumption that π is M-independent, from Theorem 5.4.6 we deduce that (D′, M, S, π) has a matroid-based packing of arborescences whose set of arcs is A′. Therefore x is the characteristic vector of this packing. The theorem follows. 5.4.4 Algorithmic aspects In this section we assume that a matroid is given by an oracle for the rank function, that is, when we give the matroid oracle a subset X ∈ S, it will give back the rank of X. The following theorem is a corollary of Theorems 2.4.1 and 5.4.6. 106 Chapter 5. Inductive constructions and decompositions Theorem 5.4.12. Let (D, M, S, π) be a matroid-based rooted-digraph. A matroid- based packing of rooted-arborescences in (D, M, S, π) or a vertex v certifying that π is not M-independent or a vertex set X certifying that (D, M, S, π) is not rooted- connected can be found in polynomial time. Proof. By the submodularity of ρD(X) + rM(SX ), Theorem 2.4.1, using the oracle on M and Theorem 5.4.6, we can either find a set violating the condition of rooted- connectedness or a vertex certifying that π is not M-independent or certify that there exists a matroid-based packing of rooted-arborescences. In the latter case, a matroid-based packing of rooted-arborescences can be found in polynomial time following the proof of Theorem 5.4.6. Using the oracle, one can test whether an arc is bad or good. When an arc uv is good, for each ′ ′ ′ ′ s ∈ Su \ cl(Sv), determine in polynomial time whether (D , M , S , π ) is rooted- ′ ′ ′ connected using the submodularity of ρD (X) + rM (SX ), the oracle for the rank function rM′ (that is easily computed from rM) and Theorem 2.4.1. The proof of Theorem 2.4.1 shows that either all arcs are bad or we find a good arc uv and ′ ′ ′ ′ s ∈ Su \ cl(Sv) such that (D , M , S , π ) is rooted-connected. In the first case, {(v, s): v ∈ V, s ∈ Sv} is the required packing. In the second case, it leads to the computation of a matroid-based packing of rooted-arborescences in (D′, M′, S′, π′) where D′ contains less arcs than D. − By the submodularity of x(RD(X)) + rM(SX ) and by Theorem 2.4.1, PM,D can be separated in polynomial time. It is a well-known result by Gr¨otschel, Lov´asz and Schrijver [49] that the separation problem and optimization problem are polynomial-time equivalent. Therefore, Theorem 5.4.12 leads to the following consequence. Theorem 5.4.13. Let (D, M, S, π) be a matroid-based rooted-digraph and c a cost function on the set of arcs of D. If there exists a matroid-based packing of rooted- arborescences in (D, M, S, π) then one of minimum cost can be found in polynomial time. 5.4.5 Further remarks Theorem 5.4.6 motivates the following extension. Given a matroid-based rooted- digraph (D, M, S, π) where M has rank function rM and a bound b : V → Z, an (M, b)-packing of rooted-arborescences is a set {(T1, s1),..., (T|S|, s|S|)} of pairwise arc-disjoint rooted-arborescences such that rM({si ∈ S : v ∈ V (Ti)}) ≥ b(v) for 107 5.4. Packing of matroid-based arborescences all v ∈ V . What is the necessary and sufficient condition for the existance of an (M, b)-packing of rooted-arborescences in (D, M, S, π)? When the function b is constant, i.e., b(v) = b for all v ∈ V , by truncating the matroid M to a matroid of rank b, then using Theorem 5.4.6, one can derive a characterization of matroid- based rooted-digraphs admitting an (M, b)-packing of rooted-arborescences. On the other hand, for general b, the problem turns out to be NP-complete since it contains the disjoint Steiner arborescences problem, that is to find 2 arc-disjoint r-arborescences both covering a specified subset of vertices. Basing on the same proof technique, Csaba Kir´aly recently extends our result to maximal independent packing of arborescences. For a subset X ⊆ V let P (X) denote the set of vertices u such that there is a directed path from u to a vertex in X, note that P (X) contains X is as a subset. When X = {v} we simply write P (v). A maximal independent packing of arborescences of (D, M, S, π) is a set {(T1, s1),..., (T|S|, s|S|)} of pairwise arc-disjoint S-rooted arborescences of D such that, for each v ∈ V , the set {si ∈ S : v ∈ V (Ti)} is independent in M and of size rM(SP (v)). Theorem 5.4.14 (Cs. Kir´aly [74]). A matroid-based rooted-digraph (D, M, S, π) has a maximal independent packing of arborescences if and only if π is M-independent and ρD(X) ≥ rM(SP (X)) − rM(SX ) holds for each non-empty X ⊆ V . This result also generalizes the result of Kamiyama, Katoh and Takizawa (The- orem 5.4.4). Moreover, Cs. Kir´aly also pointed out that Theorem 5.4.5 in turn can be obtained easily from Theorem 5.4.4 and therefore these two theorems are infact equivalent. 108 Chapter 6 Infinitesimal rigidity Contents 6.1 Introduction ...... 110 6.2 Body-bar frameworks with boundary ...... 112 6.2.1 Body-bar frameworks ...... 112 6.2.2 Body-bar frameworks with bar-boundary ...... 115 6.3 Body-length-direction frameworks ...... 119 6.3.1 Introduction ...... 119 6.3.2 Frameworks with length-direction-rigid bodies ...... 120 6.3.3 Frameworks with direction-rigid bodies ...... 122 6.3.4 Frameworks with length-rigid bodies ...... 126 6.4 Direction-length frameworks ...... 131 6.4.1 Introduction ...... 131 6.4.2 0-extensions for direction-length graphs ...... 133 6.4.3 1-extensions for direction-length graphs ...... 135 6.4.4 Union of two spanning trees ...... 141 109 6.1. Introduction 6.1 Introduction Infinitesimal rigidity plays a central role in rigidity theory. One important reason is that it is more tractable than other rigidity properties. Given a framework, its infinitesimal rigidity can be determined by calculating the rank of a rigidity matrix while determining whether the framework is locally rigid, globally rigid or universally rigid seems difficult [1, 91]. Moreover, infinitesimal rigidity implies local rigidity in general, and when we restrict ourselves to (linearly) generic frame- works, infinitesimal rigidity coincides with local rigidity [9, 65], note that almost all frameworks are (linearly) generic. Infinitesimal rigidity is also highly involved in the study of other rigidity properties. For example, Theorem 3.2.6 states that a stress matrix of maximal rank is a certificate for the global rigidity of an alge- braically generic bar-joint framework. But in most cases, it is not easy to find this certificate for a generic framework. A result of Connelly and Whiteley [25] shows that if (G, p) (not necessarily generic) is infinitesimally rigid in Rd and has a stress matrix of maximal rank then G is generically globally rigid in Rd. This property is extremely useful in proving the global rigidity preservingness of certain operations on graphs. Furthermore, infinitesimal rigidity is also important in the neighborhood stability of other rigidity properties such as globally linkedness of pairs of vertices [63]. This chapter contains results on infinitesimal rigidity of frameworks with mixed constraints. In Section 6.2 we begin by reviewing body-bar frameworks, the first model where a complete characterization of infinitesimal rigidity of generic frame- works is obtained. We then consider body-bar frameworks with bar-boundary, i.e, frameworks where some bodies are linked to an external enviroment with bars. We re-obtain a characterization of infinitesimal rigidity of generic frameworks of this type using our decomposition of graded tight graphs in Section 5.2. Section 6.3 introduces the study of body-length-direction frameworks, where several types of bodies that allow different types of motions are considered. In addition, beside distance constraints, these frameworks are also subject to direc- tion constraints. Characterization of infinitesimal rigidity for different types of frameworks are obtained. Section 6.4 considers the so-called direction-length frameworks, an extended model of bar-joint frameworks, where we concern also with direction constraints between vertices. Characterizing infinitesimal rigidity and global rigidity of these frameworks has application in Computer-Aided-Design and network localization. 110 Chapter 6. Infinitesimal rigidity Complete characterization is obtained by B. Servatius and Whiteley for infinitesi- mal rigidity in dimension 2 [94] and partial results for global rigidity are obtained by Jackson and Jord´an [61]. We introduce extension operations for these frame- works and investigate the infinitesimal rigidity preservingness of these operations in general dimension. Chapter 7 deals with the global rigidity preservingness of these operations. 111 6.2. Body-bar frameworks with boundary b a c Figure 6.1: A body-bar structure and its corresponding bar-joint graph. 6.2 Body-bar frameworks and body-bar frame- works with bar-boundary 6.2.1 Body-bar frameworks A body-bar structure is a structure constituting of solid bodies kept together by solid bars. The bodies allow translations and rotations as their motions. A body- bar structure can be modeled by a body-bar framework which is a pair (G, p) where G is a (multi)graph without loops on the body set and p is an embedding of the bars. This framework can be converted to a bar-joint framework by converting each extremity of bars to a joint and adding a complete graph on the set of joints on the same body. Figure 6.1 illustrates a body-bar structure and the corresponding bar- joint graph. However, while the problem of finding a combinatorial characterization for underlying graphs of infinitesimally rigid generic bar-joint frameworks is one central open problem in rigidity theory, the same problem for body-bar frameworks is much easier. Let G = (V,E) be a graph without loops. A realization of G as a body-bar framework is a map p which associates each edge e = uv ∈ E with a pair of points e e d 1 pu, pv ∈ R . The actual shapes and positions of bodies in V are not of significance . We may think as that each vertex u ∈ V is assigned a d-dimensional body Bu and e pu ∈ Bu for all edges e incident to u. An infinitesimal motion of a body-bar framework (G, p) is a map q that assigns to each body u ∈ V a pair (Au, tu) where Au is a d × d skew-symmetric real matrix d and tu a vector in R such that, for each edge e = uv ∈ E e e e e h(Aupu + tu) − (Avpv + tv) , pu − pvi = 0 1In fact, in bar-joint frameworks, for the purpose of studying infinitesimal rigidity, only the e e line passing through pu, pv is of significance, but not these two points themselves. 112 Chapter 6. Infinitesimal rigidity (c.f. Section 2.6). d For each u ∈ V , let wu be the 2 -dimensional vector associated with the skew- symmetric matrix Au as defined in Section 2.5. Using equation (2.5.1) we can rewrite this equality as e e e e e e e e h(pu − pv) ∨ pu , wui − h(pu − pv) ∨ pv , wvi + hpu − pv , tu − tvi = 0. e e e e e e e e e e Note that (pu − pv) ∨ pu = pu ∨ pv and (pu − pv) ∨ pv = pu ∨ pv, so the equality becomes e e e e hpu ∨ pv , wu − wvi + hpu − pv , tu − tvi = 0. For p, p′ ∈ Rd let T (p, p′) = (p, 1)∨(p′, 1) ∈ RD where (p, 1), (p′, 1) denote the (d+1)- dimensional vectors obtained by adding 1 to p, p′ as the (d + 1)-th coordinates and d+1 D = 2 . Put qu = (wu, tu). The above equality can be rewritten as